Merge branch 'master' into worksplit-multigpu

* [feat] Add ImageStitch node for concatenating images with borders

Add ImageStitch node that concatenates images in four directions with optional borders and intelligent size handling. Features include optional second image input, configurable borders with color selection, automatic batch size matching, and dimension alignment via padding or resizing.

Upstreamed from https://github.com/kijai/ComfyUI-KJNodes with enhancements for better error handling and comprehensive test coverage.

* [fix] Fix CI issues with CUDA dependencies and linting

- Mock CUDA-dependent modules in tests to avoid CI failures on CPU-only runners
- Fix ruff linting issues for code style compliance

* [fix] Improve CI compatibility by mocking nodes module import

Prevent CUDA initialization chain by mocking the nodes module at import time,
which is cleaner than deep mocking of CUDA-specific functions.

* [refactor] Clean up ImageStitch tests

- Remove unnecessary sys.path manipulation (pythonpath set in pytest.ini)
- Remove metadata tests that test framework internals rather than functionality
- Rename complex scenario test to be more descriptive of what it tests

* [refactor] Rename 'border' to 'spacing' for semantic accuracy

- Change border_width/border_color to spacing_width/spacing_color in API
- Update all tests to use spacing terminology
- Update comments and variable names throughout
- More accurately describes the gap/separator between images

CODEOWNERS

@@ -5,20 +5,20 @@
 # Inlined the team members for now.
 
 # Maintainers
-*.md @yoland68 @robinjhuang @huchenlei @webfiltered @pythongosssss @ltdrdata @Kosinkadink @christian-byrne
-/tests/ @yoland68 @robinjhuang @huchenlei @webfiltered @pythongosssss @ltdrdata @Kosinkadink @christian-byrne
-/tests-unit/ @yoland68 @robinjhuang @huchenlei @webfiltered @pythongosssss @ltdrdata @Kosinkadink @christian-byrne
-/notebooks/ @yoland68 @robinjhuang @huchenlei @webfiltered @pythongosssss @ltdrdata @Kosinkadink @christian-byrne
-/script_examples/ @yoland68 @robinjhuang @huchenlei @webfiltered @pythongosssss @ltdrdata @Kosinkadink @christian-byrne
-/.github/ @yoland68 @robinjhuang @huchenlei @webfiltered @pythongosssss @ltdrdata @Kosinkadink @christian-byrne
-/requirements.txt @yoland68 @robinjhuang @huchenlei @webfiltered @pythongosssss @ltdrdata @Kosinkadink @christian-byrne
-/pyproject.toml @yoland68 @robinjhuang @huchenlei @webfiltered @pythongosssss @ltdrdata @Kosinkadink @christian-byrne
+*.md @yoland68 @robinjhuang @webfiltered @pythongosssss @ltdrdata @Kosinkadink @christian-byrne
+/tests/ @yoland68 @robinjhuang @webfiltered @pythongosssss @ltdrdata @Kosinkadink @christian-byrne
+/tests-unit/ @yoland68 @robinjhuang @webfiltered @pythongosssss @ltdrdata @Kosinkadink @christian-byrne
+/notebooks/ @yoland68 @robinjhuang @webfiltered @pythongosssss @ltdrdata @Kosinkadink @christian-byrne
+/script_examples/ @yoland68 @robinjhuang @webfiltered @pythongosssss @ltdrdata @Kosinkadink @christian-byrne
+/.github/ @yoland68 @robinjhuang @webfiltered @pythongosssss @ltdrdata @Kosinkadink @christian-byrne
+/requirements.txt @yoland68 @robinjhuang @webfiltered @pythongosssss @ltdrdata @Kosinkadink @christian-byrne
+/pyproject.toml @yoland68 @robinjhuang @webfiltered @pythongosssss @ltdrdata @Kosinkadink @christian-byrne
 
 # Python web server
-/api_server/ @yoland68 @robinjhuang @huchenlei @webfiltered @pythongosssss @ltdrdata @christian-byrne
-/app/ @yoland68 @robinjhuang @huchenlei @webfiltered @pythongosssss @ltdrdata @christian-byrne
-/utils/ @yoland68 @robinjhuang @huchenlei @webfiltered @pythongosssss @ltdrdata @christian-byrne
+/api_server/ @yoland68 @robinjhuang @webfiltered @pythongosssss @ltdrdata @christian-byrne
+/app/ @yoland68 @robinjhuang @webfiltered @pythongosssss @ltdrdata @christian-byrne
+/utils/ @yoland68 @robinjhuang @webfiltered @pythongosssss @ltdrdata @christian-byrne
 
 # Node developers
-/comfy_extras/ @yoland68 @robinjhuang @huchenlei @pythongosssss @ltdrdata @Kosinkadink @webfiltered @christian-byrne
-/comfy/comfy_types/ @yoland68 @robinjhuang @huchenlei @pythongosssss @ltdrdata @Kosinkadink @webfiltered @christian-byrne
+/comfy_extras/ @yoland68 @robinjhuang @pythongosssss @ltdrdata @Kosinkadink @webfiltered @christian-byrne
+/comfy/comfy_types/ @yoland68 @robinjhuang @pythongosssss @ltdrdata @Kosinkadink @webfiltered @christian-byrne

app/frontend_management.py

@@ -205,6 +205,19 @@ comfyui-workflow-templates is not installed.
 """.strip()
             )
 
+    @classmethod
+    def embedded_docs_path(cls) -> str:
+        """Get the path to embedded documentation"""
+        try:
+            import comfyui_embedded_docs
+
+            return str(
+                importlib.resources.files(comfyui_embedded_docs) / "docs"
+            )
+        except ImportError:
+            logging.info("comfyui-embedded-docs package not found")
+            return None
+
     @classmethod
     def parse_version_string(cls, value: str) -> tuple[str, str, str]:
         """

comfy/cli_args.py

@@ -49,7 +49,7 @@ parser.add_argument("--temp-directory", type=str, default=None, help="Set the Co
 parser.add_argument("--input-directory", type=str, default=None, help="Set the ComfyUI input directory. Overrides --base-directory.")
 parser.add_argument("--auto-launch", action="store_true", help="Automatically launch ComfyUI in the default browser.")
 parser.add_argument("--disable-auto-launch", action="store_true", help="Disable auto launching the browser.")
-parser.add_argument("--cuda-device", type=int, default=None, metavar="DEVICE_ID", help="Set the id of the cuda device this instance will use.")
+parser.add_argument("--cuda-device", type=str, default=None, metavar="DEVICE_ID", help="Set the ids of cuda devices this instance will use.")
 cm_group = parser.add_mutually_exclusive_group()
 cm_group.add_argument("--cuda-malloc", action="store_true", help="Enable cudaMallocAsync (enabled by default for torch 2.0 and up).")
 cm_group.add_argument("--disable-cuda-malloc", action="store_true", help="Disable cudaMallocAsync.")

comfy/conds.py

@@ -24,6 +24,10 @@ class CONDRegular:
             conds.append(x.cond)
         return torch.cat(conds)
 
+    def size(self):
+        return list(self.cond.size())
+
+
 class CONDNoiseShape(CONDRegular):
     def process_cond(self, batch_size, device, area, **kwargs):
         data = self.cond
@@ -64,6 +68,7 @@ class CONDCrossAttn(CONDRegular):
             out.append(c)
         return torch.cat(out)
 
+
 class CONDConstant(CONDRegular):
     def __init__(self, cond):
         self.cond = cond
@@ -78,3 +83,48 @@ class CONDConstant(CONDRegular):
 
     def concat(self, others):
         return self.cond
+
+    def size(self):
+        return [1]
+
+
+class CONDList(CONDRegular):
+    def __init__(self, cond):
+        self.cond = cond
+
+    def process_cond(self, batch_size, device, **kwargs):
+        out = []
+        for c in self.cond:
+            out.append(comfy.utils.repeat_to_batch_size(c, batch_size).to(device))
+
+        return self._copy_with(out)
+
+    def can_concat(self, other):
+        if len(self.cond) != len(other.cond):
+            return False
+        for i in range(len(self.cond)):
+            if self.cond[i].shape != other.cond[i].shape:
+                return False
+
+        return True
+
+    def concat(self, others):
+        out = []
+        for i in range(len(self.cond)):
+            o = [self.cond[i]]
+            for x in others:
+                o.append(x.cond[i])
+            out.append(torch.cat(o))
+
+        return out
+
+    def size(self):  # hackish implementation to make the mem estimation work
+        o = 0
+        c = 1
+        for c in self.cond:
+            size = c.size()
+            o += math.prod(size)
+            if len(size) > 1:
+                c = size[1]
+
+        return [1, c, o // c]

comfy/controlnet.py

@@ -15,13 +15,14 @@
     You should have received a copy of the GNU General Public License
     along with this program.  If not, see <https://www.gnu.org/licenses/>.
 """
-
+from __future__ import annotations
 
 import torch
 from enum import Enum
 import math
 import os
 import logging
+import copy
 import comfy.utils
 import comfy.model_management
 import comfy.model_detection
@@ -36,7 +37,7 @@ import comfy.cldm.mmdit
 import comfy.ldm.hydit.controlnet
 import comfy.ldm.flux.controlnet
 import comfy.cldm.dit_embedder
-from typing import TYPE_CHECKING
+from typing import TYPE_CHECKING, Union
 if TYPE_CHECKING:
     from comfy.hooks import HookGroup
 
@@ -63,6 +64,18 @@ class StrengthType(Enum):
     CONSTANT = 1
     LINEAR_UP = 2
 
+class ControlIsolation:
+    '''Temporarily set a ControlBase object's previous_controlnet to None to prevent cascading calls.'''
+    def __init__(self, control: ControlBase):
+        self.control = control
+        self.orig_previous_controlnet = control.previous_controlnet
+
+    def __enter__(self):
+        self.control.previous_controlnet = None
+
+    def __exit__(self, *args):
+        self.control.previous_controlnet = self.orig_previous_controlnet
+
 class ControlBase:
     def __init__(self):
         self.cond_hint_original = None
@@ -76,7 +89,7 @@ class ControlBase:
         self.compression_ratio = 8
         self.upscale_algorithm = 'nearest-exact'
         self.extra_args = {}
-        self.previous_controlnet = None
+        self.previous_controlnet: Union[ControlBase, None] = None
         self.extra_conds = []
         self.strength_type = StrengthType.CONSTANT
         self.concat_mask = False
@@ -84,6 +97,7 @@ class ControlBase:
         self.extra_concat = None
         self.extra_hooks: HookGroup = None
         self.preprocess_image = lambda a: a
+        self.multigpu_clones: dict[torch.device, ControlBase] = {}
 
     def set_cond_hint(self, cond_hint, strength=1.0, timestep_percent_range=(0.0, 1.0), vae=None, extra_concat=[]):
         self.cond_hint_original = cond_hint
@@ -110,17 +124,38 @@ class ControlBase:
     def cleanup(self):
         if self.previous_controlnet is not None:
             self.previous_controlnet.cleanup()
-
+        for device_cnet in self.multigpu_clones.values():
+            with ControlIsolation(device_cnet):
+                device_cnet.cleanup()
         self.cond_hint = None
         self.extra_concat = None
         self.timestep_range = None
 
     def get_models(self):
         out = []
+        for device_cnet in self.multigpu_clones.values():
+            out += device_cnet.get_models_only_self()
         if self.previous_controlnet is not None:
             out += self.previous_controlnet.get_models()
         return out
 
+    def get_models_only_self(self):
+        'Calls get_models, but temporarily sets previous_controlnet to None.'
+        with ControlIsolation(self):
+            return self.get_models()
+
+    def get_instance_for_device(self, device):
+        'Returns instance of this Control object intended for selected device.'
+        return self.multigpu_clones.get(device, self)
+
+    def deepclone_multigpu(self, load_device, autoregister=False):
+        '''
+        Create deep clone of Control object where model(s) is set to other devices.
+
+        When autoregister is set to True, the deep clone is also added to multigpu_clones dict.
+        '''
+        raise NotImplementedError("Classes inheriting from ControlBase should define their own deepclone_multigpu funtion.")
+
     def get_extra_hooks(self):
         out = []
         if self.extra_hooks is not None:
@@ -129,7 +164,7 @@ class ControlBase:
             out += self.previous_controlnet.get_extra_hooks()
         return out
 
-    def copy_to(self, c):
+    def copy_to(self, c: ControlBase):
         c.cond_hint_original = self.cond_hint_original
         c.strength = self.strength
         c.timestep_percent_range = self.timestep_percent_range
@@ -280,6 +315,14 @@ class ControlNet(ControlBase):
         self.copy_to(c)
         return c
 
+    def deepclone_multigpu(self, load_device, autoregister=False):
+        c = self.copy()
+        c.control_model = copy.deepcopy(c.control_model)
+        c.control_model_wrapped = comfy.model_patcher.ModelPatcher(c.control_model, load_device=load_device, offload_device=comfy.model_management.unet_offload_device())
+        if autoregister:
+            self.multigpu_clones[load_device] = c
+        return c
+
     def get_models(self):
         out = super().get_models()
         out.append(self.control_model_wrapped)
@@ -805,6 +848,14 @@ class T2IAdapter(ControlBase):
         self.copy_to(c)
         return c
 
+    def deepclone_multigpu(self, load_device, autoregister=False):
+        c = self.copy()
+        c.t2i_model = copy.deepcopy(c.t2i_model)
+        c.device = load_device
+        if autoregister:
+            self.multigpu_clones[load_device] = c
+        return c
+
 def load_t2i_adapter(t2i_data, model_options={}): #TODO: model_options
     compression_ratio = 8
     upscale_algorithm = 'nearest-exact'

comfy/ldm/chroma/layers.py

@@ -80,15 +80,13 @@ class DoubleStreamBlock(nn.Module):
         (img_mod1, img_mod2), (txt_mod1, txt_mod2) = vec
 
         # prepare image for attention
-        img_modulated = self.img_norm1(img)
-        img_modulated = (1 + img_mod1.scale) * img_modulated + img_mod1.shift
+        img_modulated = torch.addcmul(img_mod1.shift, 1 + img_mod1.scale, self.img_norm1(img))
         img_qkv = self.img_attn.qkv(img_modulated)
         img_q, img_k, img_v = img_qkv.view(img_qkv.shape[0], img_qkv.shape[1], 3, self.num_heads, -1).permute(2, 0, 3, 1, 4)
         img_q, img_k = self.img_attn.norm(img_q, img_k, img_v)
 
         # prepare txt for attention
-        txt_modulated = self.txt_norm1(txt)
-        txt_modulated = (1 + txt_mod1.scale) * txt_modulated + txt_mod1.shift
+        txt_modulated = torch.addcmul(txt_mod1.shift, 1 + txt_mod1.scale, self.txt_norm1(txt))
         txt_qkv = self.txt_attn.qkv(txt_modulated)
         txt_q, txt_k, txt_v = txt_qkv.view(txt_qkv.shape[0], txt_qkv.shape[1], 3, self.num_heads, -1).permute(2, 0, 3, 1, 4)
         txt_q, txt_k = self.txt_attn.norm(txt_q, txt_k, txt_v)
@@ -102,12 +100,12 @@ class DoubleStreamBlock(nn.Module):
         txt_attn, img_attn = attn[:, : txt.shape[1]], attn[:, txt.shape[1] :]
 
         # calculate the img bloks
-        img = img + img_mod1.gate * self.img_attn.proj(img_attn)
-        img = img + img_mod2.gate * self.img_mlp((1 + img_mod2.scale) * self.img_norm2(img) + img_mod2.shift)
+        img.addcmul_(img_mod1.gate, self.img_attn.proj(img_attn))
+        img.addcmul_(img_mod2.gate, self.img_mlp(torch.addcmul(img_mod2.shift, 1 + img_mod2.scale, self.img_norm2(img))))
 
         # calculate the txt bloks
-        txt += txt_mod1.gate * self.txt_attn.proj(txt_attn)
-        txt += txt_mod2.gate * self.txt_mlp((1 + txt_mod2.scale) * self.txt_norm2(txt) + txt_mod2.shift)
+        txt.addcmul_(txt_mod1.gate, self.txt_attn.proj(txt_attn))
+        txt.addcmul_(txt_mod2.gate, self.txt_mlp(torch.addcmul(txt_mod2.shift, 1 + txt_mod2.scale, self.txt_norm2(txt))))
 
         if txt.dtype == torch.float16:
             txt = torch.nan_to_num(txt, nan=0.0, posinf=65504, neginf=-65504)
@@ -152,7 +150,7 @@ class SingleStreamBlock(nn.Module):
 
     def forward(self, x: Tensor, pe: Tensor, vec: Tensor, attn_mask=None) -> Tensor:
         mod = vec
-        x_mod = (1 + mod.scale) * self.pre_norm(x) + mod.shift
+        x_mod = torch.addcmul(mod.shift, 1 + mod.scale, self.pre_norm(x))
         qkv, mlp = torch.split(self.linear1(x_mod), [3 * self.hidden_size, self.mlp_hidden_dim], dim=-1)
 
         q, k, v = qkv.view(qkv.shape[0], qkv.shape[1], 3, self.num_heads, -1).permute(2, 0, 3, 1, 4)
@@ -162,7 +160,7 @@ class SingleStreamBlock(nn.Module):
         attn = attention(q, k, v, pe=pe, mask=attn_mask)
         # compute activation in mlp stream, cat again and run second linear layer
         output = self.linear2(torch.cat((attn, self.mlp_act(mlp)), 2))
-        x += mod.gate * output
+        x.addcmul_(mod.gate, output)
         if x.dtype == torch.float16:
             x = torch.nan_to_num(x, nan=0.0, posinf=65504, neginf=-65504)
         return x
@@ -178,6 +176,6 @@ class LastLayer(nn.Module):
         shift, scale = vec
         shift = shift.squeeze(1)
         scale = scale.squeeze(1)
-        x = (1 + scale[:, None, :]) * self.norm_final(x) + shift[:, None, :]
+        x = torch.addcmul(shift[:, None, :], 1 + scale[:, None, :], self.norm_final(x))
         x = self.linear(x)
         return x

comfy/ldm/wan/model.py

@@ -539,13 +539,20 @@ class WanModel(torch.nn.Module):
         x = self.unpatchify(x, grid_sizes)
         return x
 
-    def forward(self, x, timestep, context, clip_fea=None, transformer_options={}, **kwargs):
+    def forward(self, x, timestep, context, clip_fea=None, time_dim_concat=None, transformer_options={}, **kwargs):
         bs, c, t, h, w = x.shape
         x = comfy.ldm.common_dit.pad_to_patch_size(x, self.patch_size)
+
         patch_size = self.patch_size
         t_len = ((t + (patch_size[0] // 2)) // patch_size[0])
         h_len = ((h + (patch_size[1] // 2)) // patch_size[1])
         w_len = ((w + (patch_size[2] // 2)) // patch_size[2])
+
+        if time_dim_concat is not None:
+            time_dim_concat = comfy.ldm.common_dit.pad_to_patch_size(time_dim_concat, self.patch_size)
+            x = torch.cat([x, time_dim_concat], dim=2)
+            t_len = ((x.shape[2] + (patch_size[0] // 2)) // patch_size[0])
+
         img_ids = torch.zeros((t_len, h_len, w_len, 3), device=x.device, dtype=x.dtype)
         img_ids[:, :, :, 0] = img_ids[:, :, :, 0] + torch.linspace(0, t_len - 1, steps=t_len, device=x.device, dtype=x.dtype).reshape(-1, 1, 1)
         img_ids[:, :, :, 1] = img_ids[:, :, :, 1] + torch.linspace(0, h_len - 1, steps=h_len, device=x.device, dtype=x.dtype).reshape(1, -1, 1)

comfy/lora.py

@@ -283,8 +283,9 @@ def model_lora_keys_unet(model, key_map={}):
         for k in sdk:
             if k.startswith("diffusion_model."):
                 if k.endswith(".weight"):
-                    key_lora = k[len("diffusion_model."):-len(".weight")].replace(".", "_")
-                    key_map["lycoris_{}".format(key_lora)] = k #SimpleTuner lycoris format
+                    key_lora = k[len("diffusion_model."):-len(".weight")]
+                    key_map["lycoris_{}".format(key_lora.replace(".", "_"))] = k #SimpleTuner lycoris format
+                    key_map["transformer.{}".format(key_lora)] = k #SimpleTuner regular format
 
     if isinstance(model, comfy.model_base.ACEStep):
         for k in sdk:

comfy/model_base.py

@@ -135,6 +135,7 @@ class BaseModel(torch.nn.Module):
         logging.info("model_type {}".format(model_type.name))
         logging.debug("adm {}".format(self.adm_channels))
         self.memory_usage_factor = model_config.memory_usage_factor
+        self.memory_usage_factor_conds = ()
 
     def apply_model(self, x, t, c_concat=None, c_crossattn=None, control=None, transformer_options={}, **kwargs):
         return comfy.patcher_extension.WrapperExecutor.new_class_executor(
@@ -167,6 +168,11 @@ class BaseModel(torch.nn.Module):
             if hasattr(extra, "dtype"):
                 if extra.dtype != torch.int and extra.dtype != torch.long:
                     extra = extra.to(dtype)
+            if isinstance(extra, list):
+                ex = []
+                for ext in extra:
+                    ex.append(ext.to(dtype))
+                extra = ex
             extra_conds[o] = extra
 
         t = self.process_timestep(t, x=x, **extra_conds)
@@ -325,19 +331,28 @@ class BaseModel(torch.nn.Module):
     def scale_latent_inpaint(self, sigma, noise, latent_image, **kwargs):
         return self.model_sampling.noise_scaling(sigma.reshape([sigma.shape[0]] + [1] * (len(noise.shape) - 1)), noise, latent_image)
 
-    def memory_required(self, input_shape):
+    def memory_required(self, input_shape, cond_shapes={}):
+        input_shapes = [input_shape]
+        for c in self.memory_usage_factor_conds:
+            shape = cond_shapes.get(c, None)
+            if shape is not None and len(shape) > 0:
+                input_shapes += shape
+
         if comfy.model_management.xformers_enabled() or comfy.model_management.pytorch_attention_flash_attention():
             dtype = self.get_dtype()
             if self.manual_cast_dtype is not None:
                 dtype = self.manual_cast_dtype
             #TODO: this needs to be tweaked
-            area = input_shape[0] * math.prod(input_shape[2:])
+            area = sum(map(lambda input_shape: input_shape[0] * math.prod(input_shape[2:]), input_shapes))
             return (area * comfy.model_management.dtype_size(dtype) * 0.01 * self.memory_usage_factor) * (1024 * 1024)
         else:
             #TODO: this formula might be too aggressive since I tweaked the sub-quad and split algorithms to use less memory.
-            area = input_shape[0] * math.prod(input_shape[2:])
+            area = sum(map(lambda input_shape: input_shape[0] * math.prod(input_shape[2:]), input_shapes))
             return (area * 0.15 * self.memory_usage_factor) * (1024 * 1024)
 
+    def extra_conds_shapes(self, **kwargs):
+        return {}
+
 
 def unclip_adm(unclip_conditioning, device, noise_augmentor, noise_augment_merge=0.0, seed=None):
     adm_inputs = []
@@ -1047,6 +1062,11 @@ class WAN21(BaseModel):
         clip_vision_output = kwargs.get("clip_vision_output", None)
         if clip_vision_output is not None:
             out['clip_fea'] = comfy.conds.CONDRegular(clip_vision_output.penultimate_hidden_states)
+
+        time_dim_concat = kwargs.get("time_dim_concat", None)
+        if time_dim_concat is not None:
+            out['time_dim_concat'] = comfy.conds.CONDRegular(self.process_latent_in(time_dim_concat))
+
         return out
 
 

comfy/model_management.py

@@ -15,6 +15,7 @@
     You should have received a copy of the GNU General Public License
     along with this program.  If not, see <https://www.gnu.org/licenses/>.
 """
+from __future__ import annotations
 
 import psutil
 import logging
@@ -26,6 +27,10 @@ import platform
 import weakref
 import gc
 
+from typing import TYPE_CHECKING
+if TYPE_CHECKING:
+    from comfy.model_patcher import ModelPatcher
+
 class VRAMState(Enum):
     DISABLED = 0    #No vram present: no need to move models to vram
     NO_VRAM = 1     #Very low vram: enable all the options to save vram
@@ -171,6 +176,25 @@ def get_torch_device():
         else:
             return torch.device(torch.cuda.current_device())
 
+def get_all_torch_devices(exclude_current=False):
+    global cpu_state
+    devices = []
+    if cpu_state == CPUState.GPU:
+        if is_nvidia():
+            for i in range(torch.cuda.device_count()):
+                devices.append(torch.device(i))
+        elif is_intel_xpu():
+            for i in range(torch.xpu.device_count()):
+                devices.append(torch.device(i))
+        elif is_ascend_npu():
+            for i in range(torch.npu.device_count()):
+                devices.append(torch.device(i))
+    else:
+        devices.append(get_torch_device())
+    if exclude_current:
+        devices.remove(get_torch_device())
+    return devices
+
 def get_total_memory(dev=None, torch_total_too=False):
     global directml_enabled
     if dev is None:
@@ -297,8 +321,13 @@ except:
 
 try:
     if is_amd():
+        try:
+            rocm_version = tuple(map(int, str(torch.version.hip).split(".")[:2]))
+        except:
+            rocm_version = (6, -1)
         arch = torch.cuda.get_device_properties(get_torch_device()).gcnArchName
         logging.info("AMD arch: {}".format(arch))
+        logging.info("ROCm version: {}".format(rocm_version))
         if args.use_split_cross_attention == False and args.use_quad_cross_attention == False:
             if torch_version_numeric[0] >= 2 and torch_version_numeric[1] >= 7:  # works on 2.6 but doesn't actually seem to improve much
                 if any((a in arch) for a in ["gfx1100", "gfx1101", "gfx1151"]):  # TODO: more arches
@@ -382,9 +411,13 @@ try:
     logging.info("Device: {}".format(get_torch_device_name(get_torch_device())))
 except:
     logging.warning("Could not pick default device.")
+try:
+    for device in get_all_torch_devices(exclude_current=True):
+        logging.info("Device: {}".format(get_torch_device_name(device)))
+except:
+    pass
 
-
-current_loaded_models = []
+current_loaded_models: list[LoadedModel] = []
 
 def module_size(module):
     module_mem = 0
@@ -395,7 +428,7 @@ def module_size(module):
     return module_mem
 
 class LoadedModel:
-    def __init__(self, model):
+    def __init__(self, model: ModelPatcher):
         self._set_model(model)
         self.device = model.load_device
         self.real_model = None
@@ -403,7 +436,7 @@ class LoadedModel:
         self.model_finalizer = None
         self._patcher_finalizer = None
 
-    def _set_model(self, model):
+    def _set_model(self, model: ModelPatcher):
         self._model = weakref.ref(model)
         if model.parent is not None:
             self._parent_model = weakref.ref(model.parent)
@@ -1295,8 +1328,34 @@ def soft_empty_cache(force=False):
         torch.cuda.ipc_collect()
 
 def unload_all_models():
-    free_memory(1e30, get_torch_device())
+    for device in get_all_torch_devices():
+        free_memory(1e30, device)
 
+def unload_model_and_clones(model: ModelPatcher, unload_additional_models=True, all_devices=False):
+    'Unload only model and its clones - primarily for multigpu cloning purposes.'
+    initial_keep_loaded: list[LoadedModel] = current_loaded_models.copy()
+    additional_models = []
+    if unload_additional_models:
+        additional_models = model.get_nested_additional_models()
+    keep_loaded = []
+    for loaded_model in initial_keep_loaded:
+        if loaded_model.model is not None:
+            if model.clone_base_uuid == loaded_model.model.clone_base_uuid:
+                continue
+            # check additional models if they are a match
+            skip = False
+            for add_model in additional_models:
+                if add_model.clone_base_uuid == loaded_model.model.clone_base_uuid:
+                    skip = True
+                    break
+            if skip:
+                continue
+        keep_loaded.append(loaded_model)
+    if not all_devices:
+        free_memory(1e30, get_torch_device(), keep_loaded)
+    else:
+        for device in get_all_torch_devices():
+            free_memory(1e30, device, keep_loaded)
 
 #TODO: might be cleaner to put this somewhere else
 import threading

comfy/model_patcher.py

@@ -84,12 +84,15 @@ def set_model_options_pre_cfg_function(model_options, pre_cfg_function, disable_
 def create_model_options_clone(orig_model_options: dict):
     return comfy.patcher_extension.copy_nested_dicts(orig_model_options)
 
-def create_hook_patches_clone(orig_hook_patches):
+def create_hook_patches_clone(orig_hook_patches, copy_tuples=False):
     new_hook_patches = {}
     for hook_ref in orig_hook_patches:
         new_hook_patches[hook_ref] = {}
         for k in orig_hook_patches[hook_ref]:
             new_hook_patches[hook_ref][k] = orig_hook_patches[hook_ref][k][:]
+            if copy_tuples:
+                for i in range(len(new_hook_patches[hook_ref][k])):
+                    new_hook_patches[hook_ref][k][i] = tuple(new_hook_patches[hook_ref][k][i])
     return new_hook_patches
 
 def wipe_lowvram_weight(m):
@@ -240,6 +243,9 @@ class ModelPatcher:
         self.is_clip = False
         self.hook_mode = comfy.hooks.EnumHookMode.MaxSpeed
 
+        self.is_multigpu_base_clone = False
+        self.clone_base_uuid = uuid.uuid4()
+
         if not hasattr(self.model, 'model_loaded_weight_memory'):
             self.model.model_loaded_weight_memory = 0
 
@@ -318,18 +324,92 @@ class ModelPatcher:
         n.is_clip = self.is_clip
         n.hook_mode = self.hook_mode
 
+        n.is_multigpu_base_clone = self.is_multigpu_base_clone
+        n.clone_base_uuid = self.clone_base_uuid
+
         for callback in self.get_all_callbacks(CallbacksMP.ON_CLONE):
             callback(self, n)
         return n
 
+    def deepclone_multigpu(self, new_load_device=None, models_cache: dict[uuid.UUID,ModelPatcher]=None):
+        logging.info(f"Creating deepclone of {self.model.__class__.__name__} for {new_load_device if new_load_device else self.load_device}.")
+        comfy.model_management.unload_model_and_clones(self)
+        n = self.clone()
+        # set load device, if present
+        if new_load_device is not None:
+            n.load_device = new_load_device
+        # unlike for normal clone, backup dicts that shared same ref should not;
+        # otherwise, patchers that have deep copies of base models will erroneously influence each other.
+        n.backup = copy.deepcopy(n.backup)
+        n.object_patches_backup = copy.deepcopy(n.object_patches_backup)
+        n.hook_backup = copy.deepcopy(n.hook_backup)
+        n.model = copy.deepcopy(n.model)
+        # multigpu clone should not have multigpu additional_models entry
+        n.remove_additional_models("multigpu")
+        # multigpu_clone all stored additional_models; make sure circular references are properly handled
+        if models_cache is None:
+            models_cache = {}
+        for key, model_list in n.additional_models.items():
+            for i in range(len(model_list)):
+                add_model = n.additional_models[key][i]
+                if add_model.clone_base_uuid not in models_cache:
+                    models_cache[add_model.clone_base_uuid] = add_model.deepclone_multigpu(new_load_device=new_load_device, models_cache=models_cache)
+                n.additional_models[key][i] = models_cache[add_model.clone_base_uuid]
+        for callback in self.get_all_callbacks(CallbacksMP.ON_DEEPCLONE_MULTIGPU):
+            callback(self, n)
+        return n
+
+    def match_multigpu_clones(self):
+        multigpu_models = self.get_additional_models_with_key("multigpu")
+        if len(multigpu_models) > 0:
+            new_multigpu_models = []
+            for mm in multigpu_models:
+                # clone main model, but bring over relevant props from existing multigpu clone
+                n = self.clone()
+                n.load_device = mm.load_device
+                n.backup = mm.backup
+                n.object_patches_backup = mm.object_patches_backup
+                n.hook_backup = mm.hook_backup
+                n.model = mm.model
+                n.is_multigpu_base_clone = mm.is_multigpu_base_clone
+                n.remove_additional_models("multigpu")
+                orig_additional_models: dict[str, list[ModelPatcher]] = comfy.patcher_extension.copy_nested_dicts(n.additional_models)
+                n.additional_models = comfy.patcher_extension.copy_nested_dicts(mm.additional_models)
+                # figure out which additional models are not present in multigpu clone
+                models_cache = {}
+                for mm_add_model in mm.get_additional_models():
+                    models_cache[mm_add_model.clone_base_uuid] = mm_add_model
+                remove_models_uuids = set(list(models_cache.keys()))
+                for key, model_list in orig_additional_models.items():
+                    for orig_add_model in model_list:
+                        if orig_add_model.clone_base_uuid not in models_cache:
+                            models_cache[orig_add_model.clone_base_uuid] = orig_add_model.deepclone_multigpu(new_load_device=n.load_device, models_cache=models_cache)
+                            existing_list = n.get_additional_models_with_key(key)
+                            existing_list.append(models_cache[orig_add_model.clone_base_uuid])
+                            n.set_additional_models(key, existing_list)
+                        if orig_add_model.clone_base_uuid in remove_models_uuids:
+                            remove_models_uuids.remove(orig_add_model.clone_base_uuid)
+                # remove duplicate additional models
+                for key, model_list in n.additional_models.items():
+                    new_model_list = [x for x in model_list if x.clone_base_uuid not in remove_models_uuids]
+                    n.set_additional_models(key, new_model_list)
+                for callback in self.get_all_callbacks(CallbacksMP.ON_MATCH_MULTIGPU_CLONES):
+                    callback(self, n)
+                new_multigpu_models.append(n)
+            self.set_additional_models("multigpu", new_multigpu_models)
+
     def is_clone(self, other):
         if hasattr(other, 'model') and self.model is other.model:
             return True
         return False
 
-    def clone_has_same_weights(self, clone: 'ModelPatcher'):
-        if not self.is_clone(clone):
-            return False
+    def clone_has_same_weights(self, clone: ModelPatcher, allow_multigpu=False):
+        if allow_multigpu:
+            if self.clone_base_uuid != clone.clone_base_uuid:
+                return False
+        else:
+            if not self.is_clone(clone):
+                return False
 
         if self.current_hooks != clone.current_hooks:
             return False
@@ -929,7 +1009,7 @@ class ModelPatcher:
         return self.additional_models.get(key, [])
 
     def get_additional_models(self):
-        all_models = []
+        all_models: list[ModelPatcher] = []
         for models in self.additional_models.values():
             all_models.extend(models)
         return all_models
@@ -983,9 +1063,13 @@ class ModelPatcher:
         for callback in self.get_all_callbacks(CallbacksMP.ON_PRE_RUN):
             callback(self)
 
-    def prepare_state(self, timestep):
+    def prepare_state(self, timestep, model_options, ignore_multigpu=False):
         for callback in self.get_all_callbacks(CallbacksMP.ON_PREPARE_STATE):
-            callback(self, timestep)
+            callback(self, timestep, model_options, ignore_multigpu)
+        if not ignore_multigpu and "multigpu_clones" in model_options:
+            for p in model_options["multigpu_clones"].values():
+                p: ModelPatcher
+                p.prepare_state(timestep, model_options, ignore_multigpu=True)
 
     def restore_hook_patches(self):
         if self.hook_patches_backup is not None:
@@ -998,12 +1082,18 @@ class ModelPatcher:
     def prepare_hook_patches_current_keyframe(self, t: torch.Tensor, hook_group: comfy.hooks.HookGroup, model_options: dict[str]):
         curr_t = t[0]
         reset_current_hooks = False
+        multigpu_kf_changed_cache = None
         transformer_options = model_options.get("transformer_options", {})
         for hook in hook_group.hooks:
             changed = hook.hook_keyframe.prepare_current_keyframe(curr_t=curr_t, transformer_options=transformer_options)
             # if keyframe changed, remove any cached HookGroups that contain hook with the same hook_ref;
             # this will cause the weights to be recalculated when sampling
             if changed:
+                # cache changed for multigpu usage
+                if "multigpu_clones" in model_options:
+                    if multigpu_kf_changed_cache is None:
+                        multigpu_kf_changed_cache = []
+                    multigpu_kf_changed_cache.append(hook)
                 # reset current_hooks if contains hook that changed
                 if self.current_hooks is not None:
                     for current_hook in self.current_hooks.hooks:
@@ -1015,6 +1105,28 @@ class ModelPatcher:
                         self.cached_hook_patches.pop(cached_group)
         if reset_current_hooks:
             self.patch_hooks(None)
+        if "multigpu_clones" in model_options:
+            for p in model_options["multigpu_clones"].values():
+                p: ModelPatcher
+                p._handle_changed_hook_keyframes(multigpu_kf_changed_cache)
+
+    def _handle_changed_hook_keyframes(self, kf_changed_cache: list[comfy.hooks.Hook]):
+        'Used to handle multigpu behavior inside prepare_hook_patches_current_keyframe.'
+        if kf_changed_cache is None:
+            return
+        reset_current_hooks = False
+        # reset current_hooks if contains hook that changed
+        for hook in kf_changed_cache:
+            if self.current_hooks is not None:
+                for current_hook in self.current_hooks.hooks:
+                    if current_hook == hook:
+                        reset_current_hooks = True
+                        break
+            for cached_group in list(self.cached_hook_patches.keys()):
+                if cached_group.contains(hook):
+                    self.cached_hook_patches.pop(cached_group)
+        if reset_current_hooks:
+            self.patch_hooks(None)
 
     def register_all_hook_patches(self, hooks: comfy.hooks.HookGroup, target_dict: dict[str], model_options: dict=None,
                                   registered: comfy.hooks.HookGroup = None):

comfy/multigpu.py

@@ -0,0 +1,167 @@
+from __future__ import annotations
+import torch
+import logging
+
+from collections import namedtuple
+from typing import TYPE_CHECKING
+if TYPE_CHECKING:
+    from comfy.model_patcher import ModelPatcher
+import comfy.utils
+import comfy.patcher_extension
+import comfy.model_management
+
+
+class GPUOptions:
+    def __init__(self, device_index: int, relative_speed: float):
+        self.device_index = device_index
+        self.relative_speed = relative_speed
+
+    def clone(self):
+        return GPUOptions(self.device_index, self.relative_speed)
+
+    def create_dict(self):
+        return {
+            "relative_speed": self.relative_speed
+        }
+
+class GPUOptionsGroup:
+    def __init__(self):
+        self.options: dict[int, GPUOptions] = {}
+
+    def add(self, info: GPUOptions):
+        self.options[info.device_index] = info
+
+    def clone(self):
+        c = GPUOptionsGroup()
+        for opt in self.options.values():
+            c.add(opt)
+        return c
+
+    def register(self, model: ModelPatcher):
+        opts_dict = {}
+        # get devices that are valid for this model
+        devices: list[torch.device] = [model.load_device]
+        for extra_model in model.get_additional_models_with_key("multigpu"):
+            extra_model: ModelPatcher
+            devices.append(extra_model.load_device)
+        # create dictionary with actual device mapped to its GPUOptions
+        device_opts_list: list[GPUOptions] = []
+        for device in devices:
+            device_opts = self.options.get(device.index, GPUOptions(device_index=device.index, relative_speed=1.0))
+            opts_dict[device] = device_opts.create_dict()
+            device_opts_list.append(device_opts)
+        # make relative_speed relative to 1.0
+        min_speed = min([x.relative_speed for x in device_opts_list])
+        for value in opts_dict.values():
+            value['relative_speed'] /= min_speed
+        model.model_options['multigpu_options'] = opts_dict
+
+
+def create_multigpu_deepclones(model: ModelPatcher, max_gpus: int, gpu_options: GPUOptionsGroup=None, reuse_loaded=False):
+    'Prepare ModelPatcher to contain deepclones of its BaseModel and related properties.'
+    model = model.clone()
+    # check if multigpu is already prepared - get the load devices from them if possible to exclude
+    skip_devices = set()
+    multigpu_models = model.get_additional_models_with_key("multigpu")
+    if len(multigpu_models) > 0:
+        for mm in multigpu_models:
+            skip_devices.add(mm.load_device)
+    skip_devices = list(skip_devices)
+
+    full_extra_devices = comfy.model_management.get_all_torch_devices(exclude_current=True)
+    limit_extra_devices = full_extra_devices[:max_gpus-1]
+    extra_devices = limit_extra_devices.copy()
+    # exclude skipped devices
+    for skip in skip_devices:
+        if skip in extra_devices:
+            extra_devices.remove(skip)
+    # create new deepclones
+    if len(extra_devices) > 0:
+        for device in extra_devices:
+            device_patcher = None
+            if reuse_loaded:
+                # check if there are any ModelPatchers currently loaded that could be referenced here after a clone
+                loaded_models: list[ModelPatcher] = comfy.model_management.loaded_models()
+                for lm in loaded_models:
+                    if lm.model is not None and lm.clone_base_uuid == model.clone_base_uuid and lm.load_device == device:
+                        device_patcher = lm.clone()
+                        logging.info(f"Reusing loaded deepclone of {device_patcher.model.__class__.__name__} for {device}")
+                        break
+            if device_patcher is None:
+                device_patcher = model.deepclone_multigpu(new_load_device=device)
+                device_patcher.is_multigpu_base_clone = True
+            multigpu_models = model.get_additional_models_with_key("multigpu")
+            multigpu_models.append(device_patcher)
+            model.set_additional_models("multigpu", multigpu_models)
+        model.match_multigpu_clones()
+        if gpu_options is None:
+            gpu_options = GPUOptionsGroup()
+        gpu_options.register(model)
+    else:
+        logging.info("No extra torch devices need initialization, skipping initializing MultiGPU Work Units.")
+    # TODO: only keep model clones that don't go 'past' the intended max_gpu count
+    # multigpu_models = model.get_additional_models_with_key("multigpu")
+    # new_multigpu_models = []
+    # for m in multigpu_models:
+    #     if m.load_device in limit_extra_devices:
+    #         new_multigpu_models.append(m)
+    # model.set_additional_models("multigpu", new_multigpu_models)
+    # persist skip_devices for use in sampling code
+    # if len(skip_devices) > 0 or "multigpu_skip_devices" in model.model_options:
+    #     model.model_options["multigpu_skip_devices"] = skip_devices
+    return model
+
+
+LoadBalance = namedtuple('LoadBalance', ['work_per_device', 'idle_time'])
+def load_balance_devices(model_options: dict[str], total_work: int, return_idle_time=False, work_normalized: int=None):
+    'Optimize work assigned to different devices, accounting for their relative speeds and splittable work.'
+    opts_dict = model_options['multigpu_options']
+    devices = list(model_options['multigpu_clones'].keys())
+    speed_per_device = []
+    work_per_device = []
+    # get sum of each device's relative_speed
+    total_speed = 0.0
+    for opts in opts_dict.values():
+        total_speed += opts['relative_speed']
+    # get relative work for each device;
+    # obtained by w = (W*r)/R
+    for device in devices:
+        relative_speed = opts_dict[device]['relative_speed']
+        relative_work = (total_work*relative_speed) / total_speed
+        speed_per_device.append(relative_speed)
+        work_per_device.append(relative_work)
+    # relative work must be expressed in whole numbers, but likely is a decimal;
+    # perform rounding while maintaining total sum equal to total work (sum of relative works)
+    work_per_device = round_preserved(work_per_device)
+    dict_work_per_device = {}
+    for device, relative_work in zip(devices, work_per_device):
+        dict_work_per_device[device] = relative_work
+    if not return_idle_time:
+        return LoadBalance(dict_work_per_device, None)
+    # divide relative work by relative speed to get estimated completion time of said work by each device;
+    # time here is relative and does not correspond to real-world units
+    completion_time = [w/r for w,r in zip(work_per_device, speed_per_device)]
+    # calculate relative time spent by the devices waiting on each other after their work is completed
+    idle_time = abs(min(completion_time) - max(completion_time))
+    # if need to compare work idle time, need to normalize to a common total work
+    if work_normalized:
+        idle_time *= (work_normalized/total_work)
+
+    return LoadBalance(dict_work_per_device, idle_time)
+
+def round_preserved(values: list[float]):
+    'Round all values in a list, preserving the combined sum of values.'
+    # get floor of values; casting to int does it too
+    floored = [int(x) for x in values]
+    total_floored = sum(floored)
+    # get remainder to distribute
+    remainder = round(sum(values)) - total_floored
+    # pair values with fractional portions
+    fractional = [(i, x-floored[i]) for i, x in enumerate(values)]
+    # sort by fractional part in descending order
+    fractional.sort(key=lambda x: x[1], reverse=True)
+    # distribute the remainder
+    for i in range(remainder):
+        index = fractional[i][0]
+        floored[index] += 1
+    return floored

comfy/patcher_extension.py

@@ -3,6 +3,8 @@ from typing import Callable
 
 class CallbacksMP:
     ON_CLONE = "on_clone"
+    ON_DEEPCLONE_MULTIGPU = "on_deepclone_multigpu"
+    ON_MATCH_MULTIGPU_CLONES = "on_match_multigpu_clones"
     ON_LOAD = "on_load_after"
     ON_DETACH = "on_detach_after"
     ON_CLEANUP = "on_cleanup"

comfy/sampler_helpers.py

@@ -1,7 +1,11 @@
 from __future__ import annotations
+import torch
 import uuid
+import math
+import collections
 import comfy.model_management
 import comfy.conds
+import comfy.model_patcher
 import comfy.utils
 import comfy.hooks
 import comfy.patcher_extension
@@ -104,6 +108,62 @@ def cleanup_additional_models(models):
         if hasattr(m, 'cleanup'):
             m.cleanup()
 
+def preprocess_multigpu_conds(conds: dict[str, list[dict[str]]], model: ModelPatcher, model_options: dict[str]):
+    '''If multigpu acceleration required, creates deepclones of ControlNets and GLIGEN per device.'''
+    multigpu_models: list[ModelPatcher] = model.get_additional_models_with_key("multigpu")
+    if len(multigpu_models) == 0:
+        return
+    extra_devices = [x.load_device for x in multigpu_models]
+    # handle controlnets
+    controlnets: set[ControlBase] = set()
+    for k in conds:
+        for kk in conds[k]:
+            if 'control' in kk:
+                controlnets.add(kk['control'])
+    if len(controlnets) > 0:
+        # first, unload all controlnet clones
+        for cnet in list(controlnets):
+            cnet_models = cnet.get_models()
+            for cm in cnet_models:
+                comfy.model_management.unload_model_and_clones(cm, unload_additional_models=True)
+
+        # next, make sure each controlnet has a deepclone for all relevant devices
+        for cnet in controlnets:
+            curr_cnet = cnet
+            while curr_cnet is not None:
+                for device in extra_devices:
+                    if device not in curr_cnet.multigpu_clones:
+                        curr_cnet.deepclone_multigpu(device, autoregister=True)
+                curr_cnet = curr_cnet.previous_controlnet
+        # since all device clones are now present, recreate the linked list for cloned cnets per device
+        for cnet in controlnets:
+            curr_cnet = cnet
+            while curr_cnet is not None:
+                prev_cnet = curr_cnet.previous_controlnet
+                for device in extra_devices:
+                    device_cnet = curr_cnet.get_instance_for_device(device)
+                    prev_device_cnet = None
+                    if prev_cnet is not None:
+                        prev_device_cnet = prev_cnet.get_instance_for_device(device)
+                    device_cnet.set_previous_controlnet(prev_device_cnet)
+                curr_cnet = prev_cnet
+    # potentially handle gligen - since not widely used, ignored for now
+
+def estimate_memory(model, noise_shape, conds):
+    cond_shapes = collections.defaultdict(list)
+    cond_shapes_min = {}
+    for _, cs in conds.items():
+        for cond in cs:
+            for k, v in model.model.extra_conds_shapes(**cond).items():
+                cond_shapes[k].append(v)
+                if cond_shapes_min.get(k, None) is None:
+                    cond_shapes_min[k] = [v]
+                elif math.prod(v) > math.prod(cond_shapes_min[k][0]):
+                    cond_shapes_min[k] = [v]
+
+    memory_required = model.model.memory_required([noise_shape[0] * 2] + list(noise_shape[1:]), cond_shapes=cond_shapes)
+    minimum_memory_required = model.model.memory_required([noise_shape[0]] + list(noise_shape[1:]), cond_shapes=cond_shapes_min)
+    return memory_required, minimum_memory_required
 
 def prepare_sampling(model: ModelPatcher, noise_shape, conds, model_options=None):
     executor = comfy.patcher_extension.WrapperExecutor.new_executor(
@@ -113,13 +173,13 @@ def prepare_sampling(model: ModelPatcher, noise_shape, conds, model_options=None
     return executor.execute(model, noise_shape, conds, model_options=model_options)
 
 def _prepare_sampling(model: ModelPatcher, noise_shape, conds, model_options=None):
-    real_model: BaseModel = None
+    model.match_multigpu_clones()
+    preprocess_multigpu_conds(conds, model, model_options)
     models, inference_memory = get_additional_models(conds, model.model_dtype())
     models += get_additional_models_from_model_options(model_options)
     models += model.get_nested_additional_models()  # TODO: does this require inference_memory update?
-    memory_required = model.memory_required([noise_shape[0] * 2] + list(noise_shape[1:])) + inference_memory
-    minimum_memory_required = model.memory_required([noise_shape[0]] + list(noise_shape[1:])) + inference_memory
-    comfy.model_management.load_models_gpu([model] + models, memory_required=memory_required, minimum_memory_required=minimum_memory_required)
+    memory_required, minimum_memory_required = estimate_memory(model, noise_shape, conds)
+    comfy.model_management.load_models_gpu([model] + models, memory_required=memory_required + inference_memory, minimum_memory_required=minimum_memory_required + inference_memory)
     real_model = model.model
 
     return real_model, conds, models
@@ -133,7 +193,7 @@ def cleanup_models(conds, models):
 
     cleanup_additional_models(set(control_cleanup))
 
-def prepare_model_patcher(model: 'ModelPatcher', conds, model_options: dict):
+def prepare_model_patcher(model: ModelPatcher, conds, model_options: dict):
     '''
     Registers hooks from conds.
     '''
@@ -166,3 +226,18 @@ def prepare_model_patcher(model: 'ModelPatcher', conds, model_options: dict):
         comfy.patcher_extension.merge_nested_dicts(to_load_options.setdefault(wc_name, {}), model_options["transformer_options"][wc_name],
                                                     copy_dict1=False)
     return to_load_options
+
+def prepare_model_patcher_multigpu_clones(model_patcher: ModelPatcher, loaded_models: list[ModelPatcher], model_options: dict):
+    '''
+    In case multigpu acceleration is enabled, prep ModelPatchers for each device.
+    '''
+    multigpu_patchers: list[ModelPatcher] = [x for x in loaded_models if x.is_multigpu_base_clone]
+    if len(multigpu_patchers) > 0:
+        multigpu_dict: dict[torch.device, ModelPatcher] = {}
+        multigpu_dict[model_patcher.load_device] = model_patcher
+        for x in multigpu_patchers:
+            x.hook_patches = comfy.model_patcher.create_hook_patches_clone(model_patcher.hook_patches, copy_tuples=True)
+            x.hook_mode = model_patcher.hook_mode # match main model's hook_mode
+            multigpu_dict[x.load_device] = x
+        model_options["multigpu_clones"] = multigpu_dict
+    return multigpu_patchers

comfy/samplers.py

@@ -1,4 +1,6 @@
 from __future__ import annotations
+
+import comfy.model_management
 from .k_diffusion import sampling as k_diffusion_sampling
 from .extra_samplers import uni_pc
 from typing import TYPE_CHECKING, Callable, NamedTuple
@@ -18,6 +20,7 @@ import comfy.patcher_extension
 import comfy.hooks
 import scipy.stats
 import numpy
+import threading
 
 
 def add_area_dims(area, num_dims):
@@ -140,7 +143,7 @@ def can_concat_cond(c1, c2):
 
     return cond_equal_size(c1.conditioning, c2.conditioning)
 
-def cond_cat(c_list):
+def cond_cat(c_list, device=None):
     temp = {}
     for x in c_list:
         for k in x:
@@ -152,6 +155,8 @@ def cond_cat(c_list):
     for k in temp:
         conds = temp[k]
         out[k] = conds[0].concat(conds[1:])
+        if device is not None and hasattr(out[k], 'to'):
+            out[k] = out[k].to(device)
 
     return out
 
@@ -205,7 +210,9 @@ def calc_cond_batch(model: 'BaseModel', conds: list[list[dict]], x_in: torch.Ten
     )
     return executor.execute(model, conds, x_in, timestep, model_options)
 
-def _calc_cond_batch(model: 'BaseModel', conds: list[list[dict]], x_in: torch.Tensor, timestep, model_options):
+def _calc_cond_batch(model: 'BaseModel', conds: list[list[dict]], x_in: torch.Tensor, timestep: torch.Tensor, model_options: dict[str]):
+    if 'multigpu_clones' in model_options:
+        return _calc_cond_batch_multigpu(model, conds, x_in, timestep, model_options)
     out_conds = []
     out_counts = []
     # separate conds by matching hooks
@@ -237,7 +244,7 @@ def _calc_cond_batch(model: 'BaseModel', conds: list[list[dict]], x_in: torch.Te
     if has_default_conds:
         finalize_default_conds(model, hooked_to_run, default_conds, x_in, timestep, model_options)
 
-    model.current_patcher.prepare_state(timestep)
+    model.current_patcher.prepare_state(timestep, model_options)
 
     # run every hooked_to_run separately
     for hooks, to_run in hooked_to_run.items():
@@ -256,7 +263,13 @@ def _calc_cond_batch(model: 'BaseModel', conds: list[list[dict]], x_in: torch.Te
             for i in range(1, len(to_batch_temp) + 1):
                 batch_amount = to_batch_temp[:len(to_batch_temp)//i]
                 input_shape = [len(batch_amount) * first_shape[0]] + list(first_shape)[1:]
-                if model.memory_required(input_shape) * 1.5 < free_memory:
+                cond_shapes = collections.defaultdict(list)
+                for tt in batch_amount:
+                    cond = {k: v.size() for k, v in to_run[tt][0].conditioning.items()}
+                    for k, v in to_run[tt][0].conditioning.items():
+                        cond_shapes[k].append(v.size())
+
+                if model.memory_required(input_shape, cond_shapes=cond_shapes) * 1.5 < free_memory:
                     to_batch = batch_amount
                     break
 
@@ -339,6 +352,190 @@ def _calc_cond_batch(model: 'BaseModel', conds: list[list[dict]], x_in: torch.Te
 
     return out_conds
 
+def _calc_cond_batch_multigpu(model: BaseModel, conds: list[list[dict]], x_in: torch.Tensor, timestep: torch.Tensor, model_options: dict[str]):
+    out_conds = []
+    out_counts = []
+    # separate conds by matching hooks
+    hooked_to_run: dict[comfy.hooks.HookGroup,list[tuple[tuple,int]]] = {}
+    default_conds = []
+    has_default_conds = False
+
+    output_device = x_in.device
+
+    for i in range(len(conds)):
+        out_conds.append(torch.zeros_like(x_in))
+        out_counts.append(torch.ones_like(x_in) * 1e-37)
+
+        cond = conds[i]
+        default_c = []
+        if cond is not None:
+            for x in cond:
+                if 'default' in x:
+                    default_c.append(x)
+                    has_default_conds = True
+                    continue
+                p = get_area_and_mult(x, x_in, timestep)
+                if p is None:
+                    continue
+                if p.hooks is not None:
+                    model.current_patcher.prepare_hook_patches_current_keyframe(timestep, p.hooks, model_options)
+                hooked_to_run.setdefault(p.hooks, list())
+                hooked_to_run[p.hooks] += [(p, i)]
+        default_conds.append(default_c)
+
+    if has_default_conds:
+        finalize_default_conds(model, hooked_to_run, default_conds, x_in, timestep, model_options)
+
+    model.current_patcher.prepare_state(timestep, model_options)
+
+    devices = [dev_m for dev_m in model_options['multigpu_clones'].keys()]
+    device_batched_hooked_to_run: dict[torch.device, list[tuple[comfy.hooks.HookGroup, tuple]]] = {}
+
+    total_conds = 0
+    for to_run in hooked_to_run.values():
+        total_conds += len(to_run)
+    conds_per_device = max(1, math.ceil(total_conds//len(devices)))
+    index_device = 0
+    current_device = devices[index_device]
+    # run every hooked_to_run separately
+    for hooks, to_run in hooked_to_run.items():
+        while len(to_run) > 0:
+            current_device = devices[index_device % len(devices)]
+            batched_to_run = device_batched_hooked_to_run.setdefault(current_device, [])
+            # keep track of conds currently scheduled onto this device
+            batched_to_run_length = 0
+            for btr in batched_to_run:
+                batched_to_run_length += len(btr[1])
+
+            first = to_run[0]
+            first_shape = first[0][0].shape
+            to_batch_temp = []
+            # make sure not over conds_per_device limit when creating temp batch
+            for x in range(len(to_run)):
+                if can_concat_cond(to_run[x][0], first[0]) and len(to_batch_temp) < (conds_per_device - batched_to_run_length):
+                    to_batch_temp += [x]
+
+            to_batch_temp.reverse()
+            to_batch = to_batch_temp[:1]
+
+            free_memory = model_management.get_free_memory(current_device)
+            for i in range(1, len(to_batch_temp) + 1):
+                batch_amount = to_batch_temp[:len(to_batch_temp)//i]
+                input_shape = [len(batch_amount) * first_shape[0]] + list(first_shape)[1:]
+                if model.memory_required(input_shape) * 1.5 < free_memory:
+                    to_batch = batch_amount
+                    break
+            conds_to_batch = []
+            for x in to_batch:
+                conds_to_batch.append(to_run.pop(x))
+            batched_to_run_length += len(conds_to_batch)
+
+            batched_to_run.append((hooks, conds_to_batch))
+            if batched_to_run_length >= conds_per_device:
+                index_device += 1
+
+    thread_result = collections.namedtuple('thread_result', ['output', 'mult', 'area', 'batch_chunks', 'cond_or_uncond'])
+    def _handle_batch(device: torch.device, batch_tuple: tuple[comfy.hooks.HookGroup, tuple], results: list[thread_result]):
+        model_current: BaseModel = model_options["multigpu_clones"][device].model
+        # run every hooked_to_run separately
+        with torch.no_grad():
+            for hooks, to_batch in batch_tuple:
+                input_x = []
+                mult = []
+                c = []
+                cond_or_uncond = []
+                uuids = []
+                area = []
+                control: ControlBase = None
+                patches = None
+                for x in to_batch:
+                    o = x
+                    p = o[0]
+                    input_x.append(p.input_x)
+                    mult.append(p.mult)
+                    c.append(p.conditioning)
+                    area.append(p.area)
+                    cond_or_uncond.append(o[1])
+                    uuids.append(p.uuid)
+                    control = p.control
+                    patches = p.patches
+
+                batch_chunks = len(cond_or_uncond)
+                input_x = torch.cat(input_x).to(device)
+                c = cond_cat(c, device=device)
+                timestep_ = torch.cat([timestep.to(device)] * batch_chunks)
+
+                transformer_options = model_current.current_patcher.apply_hooks(hooks=hooks)
+                if 'transformer_options' in model_options:
+                    transformer_options = comfy.patcher_extension.merge_nested_dicts(transformer_options,
+                                                                                    model_options['transformer_options'],
+                                                                                    copy_dict1=False)
+
+                if patches is not None:
+                    # TODO: replace with merge_nested_dicts function
+                    if "patches" in transformer_options:
+                        cur_patches = transformer_options["patches"].copy()
+                        for p in patches:
+                            if p in cur_patches:
+                                cur_patches[p] = cur_patches[p] + patches[p]
+                            else:
+                                cur_patches[p] = patches[p]
+                        transformer_options["patches"] = cur_patches
+                    else:
+                        transformer_options["patches"] = patches
+
+                transformer_options["cond_or_uncond"] = cond_or_uncond[:]
+                transformer_options["uuids"] = uuids[:]
+                transformer_options["sigmas"] = timestep
+                transformer_options["sample_sigmas"] = transformer_options["sample_sigmas"].to(device)
+                transformer_options["multigpu_thread_device"] = device
+
+                cast_transformer_options(transformer_options, device=device)
+                c['transformer_options'] = transformer_options
+
+                if control is not None:
+                    device_control = control.get_instance_for_device(device)
+                    c['control'] = device_control.get_control(input_x, timestep_, c, len(cond_or_uncond), transformer_options)
+
+                if 'model_function_wrapper' in model_options:
+                    output = model_options['model_function_wrapper'](model_current.apply_model, {"input": input_x, "timestep": timestep_, "c": c, "cond_or_uncond": cond_or_uncond}).to(output_device).chunk(batch_chunks)
+                else:
+                    output = model_current.apply_model(input_x, timestep_, **c).to(output_device).chunk(batch_chunks)
+                results.append(thread_result(output, mult, area, batch_chunks, cond_or_uncond))
+
+
+    results: list[thread_result] = []
+    threads: list[threading.Thread] = []
+    for device, batch_tuple in device_batched_hooked_to_run.items():
+        new_thread = threading.Thread(target=_handle_batch, args=(device, batch_tuple, results))
+        threads.append(new_thread)
+        new_thread.start()
+
+    for thread in threads:
+        thread.join()
+
+    for output, mult, area, batch_chunks, cond_or_uncond in results:
+        for o in range(batch_chunks):
+            cond_index = cond_or_uncond[o]
+            a = area[o]
+            if a is None:
+                out_conds[cond_index] += output[o] * mult[o]
+                out_counts[cond_index] += mult[o]
+            else:
+                out_c = out_conds[cond_index]
+                out_cts = out_counts[cond_index]
+                dims = len(a) // 2
+                for i in range(dims):
+                    out_c = out_c.narrow(i + 2, a[i + dims], a[i])
+                    out_cts = out_cts.narrow(i + 2, a[i + dims], a[i])
+                out_c += output[o] * mult[o]
+                out_cts += mult[o]
+
+    for i in range(len(out_conds)):
+        out_conds[i] /= out_counts[i]
+
+    return out_conds
+
 def calc_cond_uncond_batch(model, cond, uncond, x_in, timestep, model_options): #TODO: remove
     logging.warning("WARNING: The comfy.samplers.calc_cond_uncond_batch function is deprecated please use the calc_cond_batch one instead.")
     return tuple(calc_cond_batch(model, [cond, uncond], x_in, timestep, model_options))
@@ -636,6 +833,8 @@ def pre_run_control(model, conds):
         percent_to_timestep_function = lambda a: s.percent_to_sigma(a)
         if 'control' in x:
             x['control'].pre_run(model, percent_to_timestep_function)
+            for device_cnet in x['control'].multigpu_clones.values():
+                device_cnet.pre_run(model, percent_to_timestep_function)
 
 def apply_empty_x_to_equal_area(conds, uncond, name, uncond_fill_func):
     cond_cnets = []
@@ -878,7 +1077,9 @@ def cast_to_load_options(model_options: dict[str], device=None, dtype=None):
     to_load_options = model_options.get("to_load_options", None)
     if to_load_options is None:
         return
+    cast_transformer_options(to_load_options, device, dtype)
 
+def cast_transformer_options(transformer_options: dict[str], device=None, dtype=None):
     casts = []
     if device is not None:
         casts.append(device)
@@ -887,18 +1088,17 @@ def cast_to_load_options(model_options: dict[str], device=None, dtype=None):
     # if nothing to apply, do nothing
     if len(casts) == 0:
         return
-
     # try to call .to on patches
-    if "patches" in to_load_options:
-        patches = to_load_options["patches"]
+    if "patches" in transformer_options:
+        patches = transformer_options["patches"]
         for name in patches:
             patch_list = patches[name]
             for i in range(len(patch_list)):
                 if hasattr(patch_list[i], "to"):
                     for cast in casts:
                         patch_list[i] = patch_list[i].to(cast)
-    if "patches_replace" in to_load_options:
-        patches = to_load_options["patches_replace"]
+    if "patches_replace" in transformer_options:
+        patches = transformer_options["patches_replace"]
         for name in patches:
             patch_list = patches[name]
             for k in patch_list:
@@ -908,8 +1108,8 @@ def cast_to_load_options(model_options: dict[str], device=None, dtype=None):
     # try to call .to on any wrappers/callbacks
     wrappers_and_callbacks = ["wrappers", "callbacks"]
     for wc_name in wrappers_and_callbacks:
-        if wc_name in to_load_options:
-            wc: dict[str, list] = to_load_options[wc_name]
+        if wc_name in transformer_options:
+            wc: dict[str, list] = transformer_options[wc_name]
             for wc_dict in wc.values():
                 for wc_list in wc_dict.values():
                     for i in range(len(wc_list)):
@@ -917,7 +1117,6 @@ def cast_to_load_options(model_options: dict[str], device=None, dtype=None):
                             for cast in casts:
                                 wc_list[i] = wc_list[i].to(cast)
 
-
 class CFGGuider:
     def __init__(self, model_patcher: ModelPatcher):
         self.model_patcher = model_patcher
@@ -963,6 +1162,8 @@ class CFGGuider:
         self.inner_model, self.conds, self.loaded_models = comfy.sampler_helpers.prepare_sampling(self.model_patcher, noise.shape, self.conds, self.model_options)
         device = self.model_patcher.load_device
 
+        multigpu_patchers = comfy.sampler_helpers.prepare_model_patcher_multigpu_clones(self.model_patcher, self.loaded_models, self.model_options)
+
         if denoise_mask is not None:
             denoise_mask = comfy.sampler_helpers.prepare_mask(denoise_mask, noise.shape, device)
 
@@ -973,9 +1174,13 @@ class CFGGuider:
 
         try:
             self.model_patcher.pre_run()
+            for multigpu_patcher in multigpu_patchers:
+                multigpu_patcher.pre_run()
             output = self.inner_sample(noise, latent_image, device, sampler, sigmas, denoise_mask, callback, disable_pbar, seed)
         finally:
             self.model_patcher.cleanup()
+            for multigpu_patcher in multigpu_patchers:
+                multigpu_patcher.cleanup()
 
         comfy.sampler_helpers.cleanup_models(self.conds, self.loaded_models)
         del self.inner_model

comfy/text_encoders/long_clipl.json

@@ -1,25 +0,0 @@
-{
-  "_name_or_path": "openai/clip-vit-large-patch14",
-  "architectures": [
-    "CLIPTextModel"
-  ],
-  "attention_dropout": 0.0,
-  "bos_token_id": 0,
-  "dropout": 0.0,
-  "eos_token_id": 49407,
-  "hidden_act": "quick_gelu",
-  "hidden_size": 768,
-  "initializer_factor": 1.0,
-  "initializer_range": 0.02,
-  "intermediate_size": 3072,
-  "layer_norm_eps": 1e-05,
-  "max_position_embeddings": 248,
-  "model_type": "clip_text_model",
-  "num_attention_heads": 12,
-  "num_hidden_layers": 12,
-  "pad_token_id": 1,
-  "projection_dim": 768,
-  "torch_dtype": "float32",
-  "transformers_version": "4.24.0",
-  "vocab_size": 49408
-}

comfy_api_nodes/apis/bfl_api.py

@@ -108,6 +108,24 @@ class BFLFluxProGenerateRequest(BaseModel):
     # )
 
 
+class BFLFluxKontextProGenerateRequest(BaseModel):
+    prompt: str = Field(..., description='The text prompt for what you wannt to edit.')
+    input_image: Optional[str] = Field(None, description='Image to edit in base64 format')
+    seed: Optional[int] = Field(None, description='The seed value for reproducibility.')
+    guidance: confloat(ge=0.1, le=99.0) = Field(..., description='Guidance strength for the image generation process')
+    steps: conint(ge=1, le=150) = Field(..., description='Number of steps for the image generation process')
+    safety_tolerance: Optional[conint(ge=0, le=2)] = Field(
+        2, description='Tolerance level for input and output moderation. Between 0 and 2, 0 being most strict, 6 being least strict. Defaults to 2.'
+    )
+    output_format: Optional[BFLOutputFormat] = Field(
+        BFLOutputFormat.png, description="Output format for the generated image. Can be 'jpeg' or 'png'.", examples=['png']
+    )
+    aspect_ratio: Optional[str] = Field(None, description='Aspect ratio of the image between 21:9 and 9:21.')
+    prompt_upsampling: Optional[bool] = Field(
+        None, description='Whether to perform upsampling on the prompt. If active, automatically modifies the prompt for more creative generation.'
+    )
+
+
 class BFLFluxProUltraGenerateRequest(BaseModel):
     prompt: str = Field(..., description='The text prompt for image generation.')
     prompt_upsampling: Optional[bool] = Field(

comfy_api_nodes/apis/client.py

@@ -327,7 +327,9 @@ class ApiClient:
             ApiServerError: If the API server is unreachable but internet is working
             Exception: For other request failures
         """
-        url = urljoin(self.base_url, path)
+        # Use urljoin but ensure path is relative to avoid absolute path behavior
+        relative_path = path.lstrip('/')
+        url = urljoin(self.base_url, relative_path)
         self.check_auth(self.auth_token, self.comfy_api_key)
         # Combine default headers with any provided headers
         request_headers = self.get_headers()

comfy_api_nodes/nodes_bfl.py

@@ -1,6 +1,6 @@
 import io
 from inspect import cleandoc
-from typing import Union
+from typing import Union, Optional
 from comfy.comfy_types.node_typing import IO, ComfyNodeABC
 from comfy_api_nodes.apis.bfl_api import (
     BFLStatus,
@@ -9,6 +9,7 @@ from comfy_api_nodes.apis.bfl_api import (
     BFLFluxCannyImageRequest,
     BFLFluxDepthImageRequest,
     BFLFluxProGenerateRequest,
+    BFLFluxKontextProGenerateRequest,
     BFLFluxProUltraGenerateRequest,
     BFLFluxProGenerateResponse,
 )
@@ -269,6 +270,158 @@ class FluxProUltraImageNode(ComfyNodeABC):
         return (output_image,)
 
 
+class FluxKontextProImageNode(ComfyNodeABC):
+    """
+    Edits images using Flux.1 Kontext [pro] via api based on prompt and aspect ratio.
+    """
+
+    MINIMUM_RATIO = 1 / 4
+    MAXIMUM_RATIO = 4 / 1
+    MINIMUM_RATIO_STR = "1:4"
+    MAXIMUM_RATIO_STR = "4:1"
+
+    @classmethod
+    def INPUT_TYPES(s):
+        return {
+            "required": {
+                "prompt": (
+                    IO.STRING,
+                    {
+                        "multiline": True,
+                        "default": "",
+                        "tooltip": "Prompt for the image generation - specify what and how to edit.",
+                    },
+                ),
+                "aspect_ratio": (
+                    IO.STRING,
+                    {
+                        "default": "16:9",
+                        "tooltip": "Aspect ratio of image; must be between 1:4 and 4:1.",
+                    },
+                ),
+                "guidance": (
+                    IO.FLOAT,
+                    {
+                        "default": 3.0,
+                        "min": 0.1,
+                        "max": 99.0,
+                        "step": 0.1,
+                        "tooltip": "Guidance strength for the image generation process"
+                    },
+                ),
+                "steps": (
+                    IO.INT,
+                    {
+                        "default": 50,
+                        "min": 1,
+                        "max": 150,
+                        "tooltip": "Number of steps for the image generation process"
+                    },
+                ),
+                "seed": (
+                    IO.INT,
+                    {
+                        "default": 1234,
+                        "min": 0,
+                        "max": 0xFFFFFFFFFFFFFFFF,
+                        "control_after_generate": True,
+                        "tooltip": "The random seed used for creating the noise.",
+                    },
+                ),
+                "prompt_upsampling": (
+                    IO.BOOLEAN,
+                    {
+                        "default": False,
+                        "tooltip": "Whether to perform upsampling on the prompt. If active, automatically modifies the prompt for more creative generation, but results are nondeterministic (same seed will not produce exactly the same result).",
+                    },
+                ),
+            },
+            "optional": {
+                "input_image": (IO.IMAGE,),
+            },
+            "hidden": {
+                "auth_token": "AUTH_TOKEN_COMFY_ORG",
+                "comfy_api_key": "API_KEY_COMFY_ORG",
+                "unique_id": "UNIQUE_ID",
+            },
+        }
+
+    @classmethod
+    def VALIDATE_INPUTS(cls, aspect_ratio: str):
+        try:
+            validate_aspect_ratio(
+                aspect_ratio,
+                minimum_ratio=cls.MINIMUM_RATIO,
+                maximum_ratio=cls.MAXIMUM_RATIO,
+                minimum_ratio_str=cls.MINIMUM_RATIO_STR,
+                maximum_ratio_str=cls.MAXIMUM_RATIO_STR,
+            )
+        except Exception as e:
+            return str(e)
+        return True
+
+    RETURN_TYPES = (IO.IMAGE,)
+    DESCRIPTION = cleandoc(__doc__ or "")  # Handle potential None value
+    FUNCTION = "api_call"
+    API_NODE = True
+    CATEGORY = "api node/image/BFL"
+
+    BFL_PATH = "/proxy/bfl/flux-kontext-pro/generate"
+
+    def api_call(
+        self,
+        prompt: str,
+        aspect_ratio: str,
+        guidance: float,
+        steps: int,
+        input_image: Optional[torch.Tensor]=None,
+        seed=0,
+        prompt_upsampling=False,
+        unique_id: Union[str, None] = None,
+        **kwargs,
+    ):
+        if input_image is None:
+            validate_string(prompt, strip_whitespace=False)
+        operation = SynchronousOperation(
+            endpoint=ApiEndpoint(
+                path=self.BFL_PATH,
+                method=HttpMethod.POST,
+                request_model=BFLFluxKontextProGenerateRequest,
+                response_model=BFLFluxProGenerateResponse,
+            ),
+            request=BFLFluxKontextProGenerateRequest(
+                prompt=prompt,
+                prompt_upsampling=prompt_upsampling,
+                guidance=round(guidance, 1),
+                steps=steps,
+                seed=seed,
+                aspect_ratio=validate_aspect_ratio(
+                    aspect_ratio,
+                    minimum_ratio=self.MINIMUM_RATIO,
+                    maximum_ratio=self.MAXIMUM_RATIO,
+                    minimum_ratio_str=self.MINIMUM_RATIO_STR,
+                    maximum_ratio_str=self.MAXIMUM_RATIO_STR,
+                ),
+                input_image=(
+                    input_image
+                    if input_image is None
+                    else convert_image_to_base64(input_image)
+                )
+            ),
+            auth_kwargs=kwargs,
+        )
+        output_image = handle_bfl_synchronous_operation(operation, node_id=unique_id)
+        return (output_image,)
+
+
+class FluxKontextMaxImageNode(FluxKontextProImageNode):
+    """
+    Edits images using Flux.1 Kontext [max] via api based on prompt and aspect ratio.
+    """
+
+    DESCRIPTION = cleandoc(__doc__ or "")
+    BFL_PATH = "/proxy/bfl/flux-kontext-max/generate"
+
 
 class FluxProImageNode(ComfyNodeABC):
     """
@@ -914,6 +1067,8 @@ class FluxProDepthNode(ComfyNodeABC):
 NODE_CLASS_MAPPINGS = {
     "FluxProUltraImageNode": FluxProUltraImageNode,
     # "FluxProImageNode": FluxProImageNode,
+    "FluxKontextProImageNode": FluxKontextProImageNode,
+    "FluxKontextMaxImageNode": FluxKontextMaxImageNode,
     "FluxProExpandNode": FluxProExpandNode,
     "FluxProFillNode": FluxProFillNode,
     "FluxProCannyNode": FluxProCannyNode,
@@ -924,6 +1079,8 @@ NODE_CLASS_MAPPINGS = {
 NODE_DISPLAY_NAME_MAPPINGS = {
     "FluxProUltraImageNode": "Flux 1.1 [pro] Ultra Image",
     # "FluxProImageNode": "Flux 1.1 [pro] Image",
+    "FluxKontextProImageNode": "Flux.1 Kontext [pro] Image",
+    "FluxKontextMaxImageNode": "Flux.1 Kontext [max] Image",
     "FluxProExpandNode": "Flux.1 Expand Image",
     "FluxProFillNode": "Flux.1 Fill Image",
     "FluxProCannyNode": "Flux.1 Canny Control Image",

comfy_api_nodes/nodes_pika.py

@@ -6,40 +6,42 @@ Pika API docs: https://pika-827374fb.mintlify.app/api-reference
 from __future__ import annotations
 
 import io
-from typing import Optional, TypeVar
 import logging
-import torch
+from typing import Optional, TypeVar
+
 import numpy as np
+import torch
+
+from comfy.comfy_types.node_typing import IO, ComfyNodeABC, InputTypeOptions
+from comfy_api.input_impl import VideoFromFile
+from comfy_api.input_impl.video_types import VideoCodec, VideoContainer, VideoInput
+from comfy_api_nodes.apinode_utils import (
+    download_url_to_video_output,
+    tensor_to_bytesio,
+)
 from comfy_api_nodes.apis import (
-    PikaBodyGenerate22T2vGenerate22T2vPost,
-    PikaGenerateResponse,
-    PikaBodyGenerate22I2vGenerate22I2vPost,
-    PikaVideoResponse,
-    PikaBodyGenerate22C2vGenerate22PikascenesPost,
     IngredientsMode,
-    PikaDurationEnum,
-    PikaResolutionEnum,
-    PikaBodyGeneratePikaffectsGeneratePikaffectsPost,
-    PikaBodyGeneratePikadditionsGeneratePikadditionsPost,
-    PikaBodyGeneratePikaswapsGeneratePikaswapsPost,
+    PikaBodyGenerate22C2vGenerate22PikascenesPost,
+    PikaBodyGenerate22I2vGenerate22I2vPost,
     PikaBodyGenerate22KeyframeGenerate22PikaframesPost,
+    PikaBodyGenerate22T2vGenerate22T2vPost,
+    PikaBodyGeneratePikadditionsGeneratePikadditionsPost,
+    PikaBodyGeneratePikaffectsGeneratePikaffectsPost,
+    PikaBodyGeneratePikaswapsGeneratePikaswapsPost,
+    PikaDurationEnum,
     Pikaffect,
+    PikaGenerateResponse,
+    PikaResolutionEnum,
+    PikaVideoResponse,
 )
 from comfy_api_nodes.apis.client import (
     ApiEndpoint,
-    HttpMethod,
-    SynchronousOperation,
-    PollingOperation,
     EmptyRequest,
-)
-from comfy_api_nodes.apinode_utils import (
-    tensor_to_bytesio,
-    download_url_to_video_output,
+    HttpMethod,
+    PollingOperation,
+    SynchronousOperation,
 )
 from comfy_api_nodes.mapper_utils import model_field_to_node_input
-from comfy_api.input_impl.video_types import VideoInput, VideoContainer, VideoCodec
-from comfy_api.input_impl import VideoFromFile
-from comfy.comfy_types.node_typing import IO, ComfyNodeABC, InputTypeOptions
 
 R = TypeVar("R")
 
@@ -204,6 +206,7 @@ class PikaImageToVideoV2_2(PikaNodeBase):
             "hidden": {
                 "auth_token": "AUTH_TOKEN_COMFY_ORG",
                 "comfy_api_key": "API_KEY_COMFY_ORG",
+                "unique_id": "UNIQUE_ID",
             },
         }
 
@@ -457,7 +460,7 @@ class PikAdditionsNode(PikaNodeBase):
             },
         }
 
-    DESCRIPTION = "Add any object or image into your video. Upload a video and specify what you’d like to add to create a seamlessly integrated result."
+    DESCRIPTION = "Add any object or image into your video. Upload a video and specify what you'd like to add to create a seamlessly integrated result."
 
     def api_call(
         self,

comfy_extras/nodes_images.py

@@ -14,6 +14,7 @@ import re
 from io import BytesIO
 from inspect import cleandoc
 import torch
+import comfy.utils
 
 from comfy.comfy_types import FileLocator
 
@@ -229,6 +230,186 @@ class SVG:
             all_svgs_list.extend(svg_item.data)
         return SVG(all_svgs_list)
 
+
+class ImageStitch:
+    """Upstreamed from https://github.com/kijai/ComfyUI-KJNodes"""
+
+    @classmethod
+    def INPUT_TYPES(s):
+        return {
+            "required": {
+                "image1": ("IMAGE",),
+                "direction": (["right", "down", "left", "up"], {"default": "right"}),
+                "match_image_size": ("BOOLEAN", {"default": True}),
+                "spacing_width": (
+                    "INT",
+                    {"default": 0, "min": 0, "max": 1024, "step": 2},
+                ),
+                "spacing_color": (
+                    ["white", "black", "red", "green", "blue"],
+                    {"default": "white"},
+                ),
+            },
+            "optional": {
+                "image2": ("IMAGE",),
+            },
+        }
+
+    RETURN_TYPES = ("IMAGE",)
+    FUNCTION = "stitch"
+    CATEGORY = "image/transform"
+    DESCRIPTION = """
+Stitches image2 to image1 in the specified direction.
+If image2 is not provided, returns image1 unchanged.
+Optional spacing can be added between images.
+"""
+
+    def stitch(
+        self,
+        image1,
+        direction,
+        match_image_size,
+        spacing_width,
+        spacing_color,
+        image2=None,
+    ):
+        if image2 is None:
+            return (image1,)
+
+        # Handle batch size differences
+        if image1.shape[0] != image2.shape[0]:
+            max_batch = max(image1.shape[0], image2.shape[0])
+            if image1.shape[0] < max_batch:
+                image1 = torch.cat(
+                    [image1, image1[-1:].repeat(max_batch - image1.shape[0], 1, 1, 1)]
+                )
+            if image2.shape[0] < max_batch:
+                image2 = torch.cat(
+                    [image2, image2[-1:].repeat(max_batch - image2.shape[0], 1, 1, 1)]
+                )
+
+        # Match image sizes if requested
+        if match_image_size:
+            h1, w1 = image1.shape[1:3]
+            h2, w2 = image2.shape[1:3]
+            aspect_ratio = w2 / h2
+
+            if direction in ["left", "right"]:
+                target_h, target_w = h1, int(h1 * aspect_ratio)
+            else:  # up, down
+                target_w, target_h = w1, int(w1 / aspect_ratio)
+
+            image2 = comfy.utils.common_upscale(
+                image2.movedim(-1, 1), target_w, target_h, "lanczos", "disabled"
+            ).movedim(1, -1)
+
+        # When not matching sizes, pad to align non-concat dimensions
+        if not match_image_size:
+            h1, w1 = image1.shape[1:3]
+            h2, w2 = image2.shape[1:3]
+
+            if direction in ["left", "right"]:
+                # For horizontal concat, pad heights to match
+                if h1 != h2:
+                    target_h = max(h1, h2)
+                    if h1 < target_h:
+                        pad_h = target_h - h1
+                        pad_top, pad_bottom = pad_h // 2, pad_h - pad_h // 2
+                        image1 = torch.nn.functional.pad(image1, (0, 0, 0, 0, pad_top, pad_bottom), mode='constant', value=0.0)
+                    if h2 < target_h:
+                        pad_h = target_h - h2
+                        pad_top, pad_bottom = pad_h // 2, pad_h - pad_h // 2
+                        image2 = torch.nn.functional.pad(image2, (0, 0, 0, 0, pad_top, pad_bottom), mode='constant', value=0.0)
+            else:  # up, down
+                # For vertical concat, pad widths to match
+                if w1 != w2:
+                    target_w = max(w1, w2)
+                    if w1 < target_w:
+                        pad_w = target_w - w1
+                        pad_left, pad_right = pad_w // 2, pad_w - pad_w // 2
+                        image1 = torch.nn.functional.pad(image1, (0, 0, pad_left, pad_right), mode='constant', value=0.0)
+                    if w2 < target_w:
+                        pad_w = target_w - w2
+                        pad_left, pad_right = pad_w // 2, pad_w - pad_w // 2
+                        image2 = torch.nn.functional.pad(image2, (0, 0, pad_left, pad_right), mode='constant', value=0.0)
+
+        # Ensure same number of channels
+        if image1.shape[-1] != image2.shape[-1]:
+            max_channels = max(image1.shape[-1], image2.shape[-1])
+            if image1.shape[-1] < max_channels:
+                image1 = torch.cat(
+                    [
+                        image1,
+                        torch.ones(
+                            *image1.shape[:-1],
+                            max_channels - image1.shape[-1],
+                            device=image1.device,
+                        ),
+                    ],
+                    dim=-1,
+                )
+            if image2.shape[-1] < max_channels:
+                image2 = torch.cat(
+                    [
+                        image2,
+                        torch.ones(
+                            *image2.shape[:-1],
+                            max_channels - image2.shape[-1],
+                            device=image2.device,
+                        ),
+                    ],
+                    dim=-1,
+                )
+
+        # Add spacing if specified
+        if spacing_width > 0:
+            spacing_width = spacing_width + (spacing_width % 2)  # Ensure even
+
+            color_map = {
+                "white": 1.0,
+                "black": 0.0,
+                "red": (1.0, 0.0, 0.0),
+                "green": (0.0, 1.0, 0.0),
+                "blue": (0.0, 0.0, 1.0),
+            }
+            color_val = color_map[spacing_color]
+
+            if direction in ["left", "right"]:
+                spacing_shape = (
+                    image1.shape[0],
+                    max(image1.shape[1], image2.shape[1]),
+                    spacing_width,
+                    image1.shape[-1],
+                )
+            else:
+                spacing_shape = (
+                    image1.shape[0],
+                    spacing_width,
+                    max(image1.shape[2], image2.shape[2]),
+                    image1.shape[-1],
+                )
+
+            spacing = torch.full(spacing_shape, 0.0, device=image1.device)
+            if isinstance(color_val, tuple):
+                for i, c in enumerate(color_val):
+                    if i < spacing.shape[-1]:
+                        spacing[..., i] = c
+                if spacing.shape[-1] == 4:  # Add alpha
+                    spacing[..., 3] = 1.0
+            else:
+                spacing[..., : min(3, spacing.shape[-1])] = color_val
+                if spacing.shape[-1] == 4:
+                    spacing[..., 3] = 1.0
+
+        # Concatenate images
+        images = [image2, image1] if direction in ["left", "up"] else [image1, image2]
+        if spacing_width > 0:
+            images.insert(1, spacing)
+
+        concat_dim = 2 if direction in ["left", "right"] else 1
+        return (torch.cat(images, dim=concat_dim),)
+
+
 class SaveSVGNode:
     """
     Save SVG files on disk.
@@ -318,4 +499,5 @@ NODE_CLASS_MAPPINGS = {
     "SaveAnimatedWEBP": SaveAnimatedWEBP,
     "SaveAnimatedPNG": SaveAnimatedPNG,
     "SaveSVGNode": SaveSVGNode,
+    "ImageStitch": ImageStitch,
 }

comfy_extras/nodes_multigpu.py

@@ -0,0 +1,86 @@
+from __future__ import annotations
+from inspect import cleandoc
+
+from typing import TYPE_CHECKING
+if TYPE_CHECKING:
+    from comfy.model_patcher import ModelPatcher
+import comfy.multigpu
+
+
+class MultiGPUWorkUnitsNode:
+    """
+    Prepares model to have sampling accelerated via splitting work units.
+
+    Should be placed after nodes that modify the model object itself, such as compile or attention-switch nodes.
+
+    Other than those exceptions, this node can be placed in any order.
+    """
+
+    NodeId = "MultiGPU_WorkUnits"
+    NodeName = "MultiGPU Work Units"
+    @classmethod
+    def INPUT_TYPES(cls):
+        return {
+            "required": {
+                "model": ("MODEL",),
+                "max_gpus" : ("INT", {"default": 8, "min": 1, "step": 1}),
+            },
+            "optional": {
+                "gpu_options": ("GPU_OPTIONS",)
+            }
+        }
+
+    RETURN_TYPES = ("MODEL",)
+    FUNCTION = "init_multigpu"
+    CATEGORY = "advanced/multigpu"
+    DESCRIPTION = cleandoc(__doc__)
+
+    def init_multigpu(self, model: ModelPatcher, max_gpus: int, gpu_options: comfy.multigpu.GPUOptionsGroup=None):
+        model = comfy.multigpu.create_multigpu_deepclones(model, max_gpus, gpu_options, reuse_loaded=True)
+        return (model,)
+
+class MultiGPUOptionsNode:
+    """
+    Select the relative speed of GPUs in the special case they have significantly different performance from one another.
+    """
+
+    NodeId = "MultiGPU_Options"
+    NodeName = "MultiGPU Options"
+    @classmethod
+    def INPUT_TYPES(cls):
+        return {
+            "required": {
+                "device_index": ("INT", {"default": 0, "min": 0, "max": 64}),
+                "relative_speed": ("FLOAT", {"default": 1.0, "min": 0.0, "step": 0.01})
+            },
+            "optional": {
+                "gpu_options": ("GPU_OPTIONS",)
+            }
+        }
+
+    RETURN_TYPES = ("GPU_OPTIONS",)
+    FUNCTION = "create_gpu_options"
+    CATEGORY = "advanced/multigpu"
+    DESCRIPTION = cleandoc(__doc__)
+
+    def create_gpu_options(self, device_index: int, relative_speed: float, gpu_options: comfy.multigpu.GPUOptionsGroup=None):
+        if not gpu_options:
+            gpu_options = comfy.multigpu.GPUOptionsGroup()
+        gpu_options.clone()
+
+        opt = comfy.multigpu.GPUOptions(device_index=device_index, relative_speed=relative_speed)
+        gpu_options.add(opt)
+
+        return (gpu_options,)
+
+
+node_list = [
+    MultiGPUWorkUnitsNode,
+    MultiGPUOptionsNode
+]
+NODE_CLASS_MAPPINGS = {}
+NODE_DISPLAY_NAME_MAPPINGS = {}
+
+for node in node_list:
+    NODE_CLASS_MAPPINGS[node.NodeId] = node
+    NODE_DISPLAY_NAME_MAPPINGS[node.NodeId] = node.NodeName

comfy_extras/nodes_string.py

@@ -296,6 +296,41 @@ class RegexExtract():
 
         return result,
 
+
+class RegexReplace():
+    DESCRIPTION = "Find and replace text using regex patterns."
+    @classmethod
+    def INPUT_TYPES(s):
+        return {
+            "required": {
+                "string": (IO.STRING, {"multiline": True}),
+                "regex_pattern": (IO.STRING, {"multiline": True}),
+                "replace": (IO.STRING, {"multiline": True}),
+            },
+            "optional": {
+                "case_insensitive": (IO.BOOLEAN, {"default": True}),
+                "multiline": (IO.BOOLEAN, {"default": False}),
+                "dotall": (IO.BOOLEAN, {"default": False, "tooltip": "When enabled, the dot (.) character will match any character including newline characters. When disabled, dots won't match newlines."}),
+                "count": (IO.INT, {"default": 0, "min": 0, "max": 100, "tooltip": "Maximum number of replacements to make. Set to 0 to replace all occurrences (default). Set to 1 to replace only the first match, 2 for the first two matches, etc."}),
+            }
+        }
+
+    RETURN_TYPES = (IO.STRING,)
+    FUNCTION = "execute"
+    CATEGORY = "utils/string"
+
+    def execute(self, string, regex_pattern, replace, case_insensitive=True, multiline=False, dotall=False, count=0, **kwargs):
+        flags = 0
+
+        if case_insensitive:
+            flags |= re.IGNORECASE
+        if multiline:
+            flags |= re.MULTILINE
+        if dotall:
+            flags |= re.DOTALL
+        result = re.sub(regex_pattern, replace, string, count=count, flags=flags)
+        return result,
+
 NODE_CLASS_MAPPINGS = {
     "StringConcatenate": StringConcatenate,
     "StringSubstring": StringSubstring,
@@ -306,7 +341,8 @@ NODE_CLASS_MAPPINGS = {
     "StringContains": StringContains,
     "StringCompare": StringCompare,
     "RegexMatch": RegexMatch,
-    "RegexExtract": RegexExtract
+    "RegexExtract": RegexExtract,
+    "RegexReplace": RegexReplace,
 }
 
 NODE_DISPLAY_NAME_MAPPINGS = {
@@ -319,5 +355,6 @@ NODE_DISPLAY_NAME_MAPPINGS = {
     "StringContains": "Contains",
     "StringCompare": "Compare",
     "RegexMatch": "Regex Match",
-    "RegexExtract": "Regex Extract"
+    "RegexExtract": "Regex Extract",
+    "RegexReplace": "Regex Replace",
 }

comfy_extras/nodes_wan.py

@@ -345,6 +345,44 @@ class WanCameraImageToVideo:
         out_latent["samples"] = latent
         return (positive, negative, out_latent)
 
+class WanPhantomSubjectToVideo:
+    @classmethod
+    def INPUT_TYPES(s):
+        return {"required": {"positive": ("CONDITIONING", ),
+                             "negative": ("CONDITIONING", ),
+                             "vae": ("VAE", ),
+                             "width": ("INT", {"default": 832, "min": 16, "max": nodes.MAX_RESOLUTION, "step": 16}),
+                             "height": ("INT", {"default": 480, "min": 16, "max": nodes.MAX_RESOLUTION, "step": 16}),
+                             "length": ("INT", {"default": 81, "min": 1, "max": nodes.MAX_RESOLUTION, "step": 4}),
+                             "batch_size": ("INT", {"default": 1, "min": 1, "max": 4096}),
+                },
+                "optional": {"images": ("IMAGE", ),
+                }}
+
+    RETURN_TYPES = ("CONDITIONING", "CONDITIONING", "CONDITIONING", "LATENT")
+    RETURN_NAMES = ("positive", "negative_text", "negative_img_text", "latent")
+    FUNCTION = "encode"
+
+    CATEGORY = "conditioning/video_models"
+
+    def encode(self, positive, negative, vae, width, height, length, batch_size, images):
+        latent = torch.zeros([batch_size, 16, ((length - 1) // 4) + 1, height // 8, width // 8], device=comfy.model_management.intermediate_device())
+        cond2 = negative
+        if images is not None:
+            images = comfy.utils.common_upscale(images[:length].movedim(-1, 1), width, height, "bilinear", "center").movedim(1, -1)
+            latent_images = []
+            for i in images:
+                latent_images += [vae.encode(i.unsqueeze(0)[:, :, :, :3])]
+            concat_latent_image = torch.cat(latent_images, dim=2)
+
+            positive = node_helpers.conditioning_set_values(positive, {"time_dim_concat": concat_latent_image})
+            cond2 = node_helpers.conditioning_set_values(negative, {"time_dim_concat": concat_latent_image})
+            negative = node_helpers.conditioning_set_values(negative, {"time_dim_concat": comfy.latent_formats.Wan21().process_out(torch.zeros_like(concat_latent_image))})
+
+        out_latent = {}
+        out_latent["samples"] = latent
+        return (positive, cond2, negative, out_latent)
+
 NODE_CLASS_MAPPINGS = {
     "WanImageToVideo": WanImageToVideo,
     "WanFunControlToVideo": WanFunControlToVideo,
@@ -353,4 +391,5 @@ NODE_CLASS_MAPPINGS = {
     "WanVaceToVideo": WanVaceToVideo,
     "TrimVideoLatent": TrimVideoLatent,
     "WanCameraImageToVideo": WanCameraImageToVideo,
+    "WanPhantomSubjectToVideo": WanPhantomSubjectToVideo,
 }

comfyui_version.py

@@ -1,3 +1,3 @@
 # This file is automatically generated by the build process when version is
 # updated in pyproject.toml.
-__version__ = "0.3.37"
+__version__ = "0.3.39"

nodes.py

@@ -2061,6 +2061,7 @@ NODE_DISPLAY_NAME_MAPPINGS = {
     "ImagePadForOutpaint": "Pad Image for Outpainting",
     "ImageBatch": "Batch Images",
     "ImageCrop": "Image Crop",
+    "ImageStitch": "Image Stitch",
     "ImageBlend": "Image Blend",
     "ImageBlur": "Image Blur",
     "ImageQuantize": "Image Quantize",
@@ -2240,6 +2241,7 @@ def init_builtin_extra_nodes():
         "nodes_mahiro.py",
         "nodes_lt.py",
         "nodes_hooks.py",
+        "nodes_multigpu.py",
         "nodes_load_3d.py",
         "nodes_cosmos.py",
         "nodes_video.py",

pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "ComfyUI"
-version = "0.3.37"
+version = "0.3.39"
 readme = "README.md"
 license = { file = "LICENSE" }
 requires-python = ">=3.9"

requirements.txt

@@ -1,5 +1,6 @@
-comfyui-frontend-package==1.20.6
-comfyui-workflow-templates==0.1.20
+comfyui-frontend-package==1.21.3
+comfyui-workflow-templates==0.1.23
+comfyui-embedded-docs==0.2.0
 torch
 torchsde
 torchvision

server.py

@@ -746,6 +746,13 @@ class PromptServer():
                 web.static('/templates', workflow_templates_path)
             ])
 
+        # Serve embedded documentation from the package
+        embedded_docs_path = FrontendManager.embedded_docs_path()
+        if embedded_docs_path:
+            self.app.add_routes([
+                web.static('/docs', embedded_docs_path)
+            ])
+
         self.app.add_routes([
             web.static('/', self.web_root),
         ])

tests-unit/comfy_extras_test/image_stitch_test.py

@@ -0,0 +1,240 @@
+import torch
+from unittest.mock import patch, MagicMock
+
+# Mock nodes module to prevent CUDA initialization during import
+mock_nodes = MagicMock()
+mock_nodes.MAX_RESOLUTION = 16384
+
+with patch.dict('sys.modules', {'nodes': mock_nodes}):
+    from comfy_extras.nodes_images import ImageStitch
+
+
+class TestImageStitch:
+
+    def create_test_image(self, batch_size=1, height=64, width=64, channels=3):
+        """Helper to create test images with specific dimensions"""
+        return torch.rand(batch_size, height, width, channels)
+
+    def test_no_image2_passthrough(self):
+        """Test that when image2 is None, image1 is returned unchanged"""
+        node = ImageStitch()
+        image1 = self.create_test_image()
+
+        result = node.stitch(image1, "right", True, 0, "white", image2=None)
+
+        assert len(result) == 1
+        assert torch.equal(result[0], image1)
+
+    def test_basic_horizontal_stitch_right(self):
+        """Test basic horizontal stitching to the right"""
+        node = ImageStitch()
+        image1 = self.create_test_image(height=32, width=32)
+        image2 = self.create_test_image(height=32, width=24)
+
+        result = node.stitch(image1, "right", False, 0, "white", image2)
+
+        assert result[0].shape == (1, 32, 56, 3)  # 32 + 24 width
+
+    def test_basic_horizontal_stitch_left(self):
+        """Test basic horizontal stitching to the left"""
+        node = ImageStitch()
+        image1 = self.create_test_image(height=32, width=32)
+        image2 = self.create_test_image(height=32, width=24)
+
+        result = node.stitch(image1, "left", False, 0, "white", image2)
+
+        assert result[0].shape == (1, 32, 56, 3)  # 24 + 32 width
+
+    def test_basic_vertical_stitch_down(self):
+        """Test basic vertical stitching downward"""
+        node = ImageStitch()
+        image1 = self.create_test_image(height=32, width=32)
+        image2 = self.create_test_image(height=24, width=32)
+
+        result = node.stitch(image1, "down", False, 0, "white", image2)
+
+        assert result[0].shape == (1, 56, 32, 3)  # 32 + 24 height
+
+    def test_basic_vertical_stitch_up(self):
+        """Test basic vertical stitching upward"""
+        node = ImageStitch()
+        image1 = self.create_test_image(height=32, width=32)
+        image2 = self.create_test_image(height=24, width=32)
+
+        result = node.stitch(image1, "up", False, 0, "white", image2)
+
+        assert result[0].shape == (1, 56, 32, 3)  # 24 + 32 height
+
+    def test_size_matching_horizontal(self):
+        """Test size matching for horizontal concatenation"""
+        node = ImageStitch()
+        image1 = self.create_test_image(height=64, width=64)
+        image2 = self.create_test_image(height=32, width=32)  # Different aspect ratio
+
+        result = node.stitch(image1, "right", True, 0, "white", image2)
+
+        # image2 should be resized to match image1's height (64) with preserved aspect ratio
+        expected_width = 64 + 64  # original + resized (32*64/32 = 64)
+        assert result[0].shape == (1, 64, expected_width, 3)
+
+    def test_size_matching_vertical(self):
+        """Test size matching for vertical concatenation"""
+        node = ImageStitch()
+        image1 = self.create_test_image(height=64, width=64)
+        image2 = self.create_test_image(height=32, width=32)
+
+        result = node.stitch(image1, "down", True, 0, "white", image2)
+
+        # image2 should be resized to match image1's width (64) with preserved aspect ratio
+        expected_height = 64 + 64  # original + resized (32*64/32 = 64)
+        assert result[0].shape == (1, expected_height, 64, 3)
+
+    def test_padding_for_mismatched_heights_horizontal(self):
+        """Test padding when heights don't match in horizontal concatenation"""
+        node = ImageStitch()
+        image1 = self.create_test_image(height=64, width=32)
+        image2 = self.create_test_image(height=48, width=24)  # Shorter height
+
+        result = node.stitch(image1, "right", False, 0, "white", image2)
+
+        # Both images should be padded to height 64
+        assert result[0].shape == (1, 64, 56, 3)  # 32 + 24 width, max(64,48) height
+
+    def test_padding_for_mismatched_widths_vertical(self):
+        """Test padding when widths don't match in vertical concatenation"""
+        node = ImageStitch()
+        image1 = self.create_test_image(height=32, width=64)
+        image2 = self.create_test_image(height=24, width=48)  # Narrower width
+
+        result = node.stitch(image1, "down", False, 0, "white", image2)
+
+        # Both images should be padded to width 64
+        assert result[0].shape == (1, 56, 64, 3)  # 32 + 24 height, max(64,48) width
+
+    def test_spacing_horizontal(self):
+        """Test spacing addition in horizontal concatenation"""
+        node = ImageStitch()
+        image1 = self.create_test_image(height=32, width=32)
+        image2 = self.create_test_image(height=32, width=24)
+        spacing_width = 16
+
+        result = node.stitch(image1, "right", False, spacing_width, "white", image2)
+
+        # Expected width: 32 + 16 (spacing) + 24 = 72
+        assert result[0].shape == (1, 32, 72, 3)
+
+    def test_spacing_vertical(self):
+        """Test spacing addition in vertical concatenation"""
+        node = ImageStitch()
+        image1 = self.create_test_image(height=32, width=32)
+        image2 = self.create_test_image(height=24, width=32)
+        spacing_width = 16
+
+        result = node.stitch(image1, "down", False, spacing_width, "white", image2)
+
+        # Expected height: 32 + 16 (spacing) + 24 = 72
+        assert result[0].shape == (1, 72, 32, 3)
+
+    def test_spacing_color_values(self):
+        """Test that spacing colors are applied correctly"""
+        node = ImageStitch()
+        image1 = self.create_test_image(height=32, width=32)
+        image2 = self.create_test_image(height=32, width=32)
+
+        # Test white spacing
+        result_white = node.stitch(image1, "right", False, 16, "white", image2)
+        # Check that spacing region contains white values (close to 1.0)
+        spacing_region = result_white[0][:, :, 32:48, :]  # Middle 16 pixels
+        assert torch.all(spacing_region >= 0.9)  # Should be close to white
+
+        # Test black spacing
+        result_black = node.stitch(image1, "right", False, 16, "black", image2)
+        spacing_region = result_black[0][:, :, 32:48, :]
+        assert torch.all(spacing_region <= 0.1)  # Should be close to black
+
+    def test_odd_spacing_width_made_even(self):
+        """Test that odd spacing widths are made even"""
+        node = ImageStitch()
+        image1 = self.create_test_image(height=32, width=32)
+        image2 = self.create_test_image(height=32, width=32)
+
+        # Use odd spacing width
+        result = node.stitch(image1, "right", False, 15, "white", image2)
+
+        # Should be made even (16), so total width = 32 + 16 + 32 = 80
+        assert result[0].shape == (1, 32, 80, 3)
+
+    def test_batch_size_matching(self):
+        """Test that different batch sizes are handled correctly"""
+        node = ImageStitch()
+        image1 = self.create_test_image(batch_size=2, height=32, width=32)
+        image2 = self.create_test_image(batch_size=1, height=32, width=32)
+
+        result = node.stitch(image1, "right", False, 0, "white", image2)
+
+        # Should match larger batch size
+        assert result[0].shape == (2, 32, 64, 3)
+
+    def test_channel_matching_rgb_to_rgba(self):
+        """Test that channel differences are handled (RGB + alpha)"""
+        node = ImageStitch()
+        image1 = self.create_test_image(channels=3)  # RGB
+        image2 = self.create_test_image(channels=4)  # RGBA
+
+        result = node.stitch(image1, "right", False, 0, "white", image2)
+
+        # Should have 4 channels (RGBA)
+        assert result[0].shape[-1] == 4
+
+    def test_channel_matching_rgba_to_rgb(self):
+        """Test that channel differences are handled (RGBA + RGB)"""
+        node = ImageStitch()
+        image1 = self.create_test_image(channels=4)  # RGBA
+        image2 = self.create_test_image(channels=3)  # RGB
+
+        result = node.stitch(image1, "right", False, 0, "white", image2)
+
+        # Should have 4 channels (RGBA)
+        assert result[0].shape[-1] == 4
+
+    def test_all_color_options(self):
+        """Test all available color options"""
+        node = ImageStitch()
+        image1 = self.create_test_image(height=32, width=32)
+        image2 = self.create_test_image(height=32, width=32)
+
+        colors = ["white", "black", "red", "green", "blue"]
+
+        for color in colors:
+            result = node.stitch(image1, "right", False, 16, color, image2)
+            assert result[0].shape == (1, 32, 80, 3)  # Basic shape check
+
+    def test_all_directions(self):
+        """Test all direction options"""
+        node = ImageStitch()
+        image1 = self.create_test_image(height=32, width=32)
+        image2 = self.create_test_image(height=32, width=32)
+
+        directions = ["right", "left", "up", "down"]
+
+        for direction in directions:
+            result = node.stitch(image1, direction, False, 0, "white", image2)
+            assert result[0].shape == (1, 32, 64, 3) if direction in ["right", "left"] else (1, 64, 32, 3)
+
+    def test_batch_size_channel_spacing_integration(self):
+        """Test integration of batch matching, channel matching, size matching, and spacings"""
+        node = ImageStitch()
+        image1 = self.create_test_image(batch_size=2, height=64, width=48, channels=3)
+        image2 = self.create_test_image(batch_size=1, height=32, width=32, channels=4)
+
+        result = node.stitch(image1, "right", True, 8, "red", image2)
+
+        # Should handle: batch matching, size matching, channel matching, spacing
+        assert result[0].shape[0] == 2  # Batch size matched
+        assert result[0].shape[-1] == 4  # Channels matched to max
+        assert result[0].shape[1] == 64  # Height from image1 (size matching)
+        # Width should be: 48 + 8 (spacing) + resized_image2_width
+        expected_image2_width = int(64 * (32/32))  # Resized to height 64
+        expected_total_width = 48 + 8 + expected_image2_width
+        assert result[0].shape[2] == expected_total_width
+