diff --git a/modules/sd_hijack.py b/modules/sd_hijack.py index 307cc67dd..f12a9696f 100644 --- a/modules/sd_hijack.py +++ b/modules/sd_hijack.py @@ -281,8 +281,15 @@ class FrozenCLIPEmbedderWithCustomWords(torch.nn.Module): remade_batch_tokens_of_same_length = [x + [self.wrapped.tokenizer.eos_token_id] * (target_token_count - len(x)) for x in remade_batch_tokens] tokens = torch.asarray(remade_batch_tokens_of_same_length).to(device) - outputs = self.wrapped.transformer(input_ids=tokens, position_ids=position_ids) - z = outputs.last_hidden_state + + tmp = -opts.CLIP_ignore_last_layers + if (opts.CLIP_ignore_last_layers == 0): + outputs = self.wrapped.transformer(input_ids=tokens, position_ids=position_ids) + z = outputs.last_hidden_state + else: + outputs = self.wrapped.transformer(input_ids=tokens, position_ids=position_ids, output_hidden_states=tmp) + z = outputs.hidden_states[tmp] + z = self.wrapped.transformer.text_model.final_layer_norm(z) # restoring original mean is likely not correct, but it seems to work well to prevent artifacts that happen otherwise batch_multipliers_of_same_length = [x + [1.0] * (target_token_count - len(x)) for x in batch_multipliers] diff --git a/modules/shared.py b/modules/shared.py index 8f9412262..af8dc7447 100644 --- a/modules/shared.py +++ b/modules/shared.py @@ -225,6 +225,7 @@ options_templates.update(options_section(('sd', "Stable Diffusion"), { "use_old_emphasis_implementation": OptionInfo(False, "Use old emphasis implementation. Can be useful to reproduce old seeds."), "enable_batch_seeds": OptionInfo(True, "Make K-diffusion samplers produce same images in a batch as when making a single image"), "filter_nsfw": OptionInfo(False, "Filter NSFW content"), + 'CLIP_ignore_last_layers': OptionInfo(0, "Ignore last layers of CLIP model", gr.Slider, {"minimum": 0, "maximum": 5, "step": 1}), "random_artist_categories": OptionInfo([], "Allowed categories for random artists selection when using the Roll button", gr.CheckboxGroup, {"choices": artist_db.categories()}), }))