ryan/ComfyUI
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Cleaner CLIP text encoder implementation.

Browse Source 
Use a simple CLIP model implementation instead of the one from
transformers.

This will allow some interesting things that would too hackish to implement
using the transformers implementation.
This commit is contained in:
comfyanonymous
2023-12-06 15:55:09 -05:00
parent 2db86b4676
commit fbdb14d4c4
5 changed files with 172 additions and 49 deletions
Download Patch File Download Diff File Expand all files Collapse all files

6
comfy/sd2_clip.py
View File

@@ -3,13 +3,13 @@ import torch
import os
class SD2ClipHModel(sd1_clip.SDClipModel):
def __init__(self, arch="ViT-H-14", device="cpu", max_length=77, freeze=True, layer="penultimate", layer_idx=None, textmodel_path=None, dtype=None):
def __init__(self, arch="ViT-H-14", device="cpu", max_length=77, freeze=True, layer="penultimate", layer_idx=None, dtype=None):
if layer == "penultimate":
layer="hidden"
layer_idx=23
layer_idx=-2
textmodel_json_config = os.path.join(os.path.dirname(os.path.realpath(__file__)), "sd2_clip_config.json")
super().__init__(device=device, freeze=freeze, layer=layer, layer_idx=layer_idx, textmodel_json_config=textmodel_json_config, textmodel_path=textmodel_path, dtype=dtype, special_tokens={"start": 49406, "end": 49407, "pad": 0})
super().__init__(device=device, freeze=freeze, layer=layer, layer_idx=layer_idx, textmodel_json_config=textmodel_json_config, dtype=dtype, special_tokens={"start": 49406, "end": 49407, "pad": 0})
class SD2ClipHTokenizer(sd1_clip.SDTokenizer):
def __init__(self, tokenizer_path=None, embedding_directory=None):