Public/automatic
1
0
Fork 0
mirror of https://github.com/vladmandic/automatic synced 2026-09-19 17:24:32 +02:00
Code Issues Packages Projects Releases Wiki Activity

fix(prompt): restore textual inversion on clip-skip>=2

Browse Source 
transformers 5.6 flattened CLIPTextModel, removing the .text_model wrapper that
compel_hijack and the xhinker parser dereference on the normalized clip-skip path.
On SD1.5 at clip-skip >= 2 this raised AttributeError, which processing_prompt
caught and silently fell back to fixed-attention encoding, dropping textual
inversion and prompt weighting.

Resolve the submodule via getattr(te, 'text_model', te), correct for flattened
CLIPTextModel, CLIPTextModelWithProjection (still nested), and transformers < 5.6.
This commit is contained in:
CalamitousFelicitousness
2026-06-25 02:06:05 +01:00
parent 7ed3cd71e0
commit aeef359bd7
2 changed files with 8 additions and 4 deletions
Download Patch File Download Diff File Expand all files Collapse all files
modules/prompt_parser_diffusers.py
+3 -1
View File
@@ -327,7 +327,9 @@ def compel_hijack(self, token_ids: torch.Tensor, attention_mask: torch.Tensor |
else:
hidden_state = text_encoder_output.hidden_states[-clip_skip]
if normalized:
hidden_state = self.text_encoder.text_model.final_layer_norm(hidden_state)
# transformers >=5.6 flattened CLIPTextModel; CLIPTextModelWithProjection still nests it under .text_model
text_model = getattr(self.text_encoder, 'text_model', self.text_encoder)
hidden_state = text_model.final_layer_norm(hidden_state)
return hidden_state
modules/prompt_parser_xhinker.py
+5 -3
View File
@@ -216,9 +216,11 @@ def get_weighted_text_embeddings_sd15(
, generator = torch.Generator(text2img_pipe.device).manual_seed(2)
).images[0]
"""
original_clip_layers = pipe.text_encoder.text_model.encoder.layers
# transformers >=5.6 flattened CLIPTextModel; CLIPTextModelWithProjection still nests it under .text_model
clip_text_model = getattr(pipe.text_encoder, 'text_model', pipe.text_encoder)
original_clip_layers = clip_text_model.encoder.layers
if clip_skip > 0:
pipe.text_encoder.text_model.encoder.layers = original_clip_layers[:-clip_skip]
clip_text_model.encoder.layers = original_clip_layers[:-clip_skip]
eos = pipe.tokenizer.eos_token_id
prompt_tokens, prompt_weights = get_prompts_tokens_with_weights(
@@ -310,7 +312,7 @@ def get_weighted_text_embeddings_sd15(
# recover clip layers
if clip_skip > 0:
pipe.text_encoder.text_model.encoder.layers = original_clip_layers
clip_text_model.encoder.layers = original_clip_layers
return prompt_embeds, neg_prompt_embeds