Merge pull request #15992 from silveroxides/dev

Add option to enable clip skip for clip L on SDXL

modules/sd_hijack_clip.py

@@ -353,7 +353,9 @@ class FrozenCLIPEmbedderForSDXLWithCustomWords(FrozenCLIPEmbedderWithCustomWords
     def encode_with_transformers(self, tokens):
         outputs = self.wrapped.transformer(input_ids=tokens, output_hidden_states=self.wrapped.layer == "hidden")
 
-        if self.wrapped.layer == "last":
+        if opts.sdxl_clip_l_skip is True:
+            z = outputs.hidden_states[-opts.CLIP_stop_at_last_layers]
+        elif self.wrapped.layer == "last":
             z = outputs.last_hidden_state
         else:
             z = outputs.hidden_states[self.wrapped.layer_idx]

modules/shared_options.py

@@ -176,6 +176,7 @@ options_templates.update(options_section(('sd', "Stable Diffusion", "sd"), {
     "emphasis": OptionInfo("Original", "Emphasis mode", gr.Radio, lambda: {"choices": [x.name for x in sd_emphasis.options]}, infotext="Emphasis").info("makes it possible to make model to pay (more:1.1) or (less:0.9) attention to text when you use the syntax in prompt; " + sd_emphasis.get_options_descriptions()),
     "enable_batch_seeds": OptionInfo(True, "Make K-diffusion samplers produce same images in a batch as when making a single image"),
     "comma_padding_backtrack": OptionInfo(20, "Prompt word wrap length limit", gr.Slider, {"minimum": 0, "maximum": 74, "step": 1}).info("in tokens - for texts shorter than specified, if they don't fit into 75 token limit, move them to the next 75 token chunk"),
+    "sdxl_clip_l_skip": OptionInfo(False, "Clip skip SDXL", gr.Checkbox).info("Enable Clip skip for the secondary clip model in sdxl. Has no effect on SD 1.5 or SD 2.0/2.1."),
     "CLIP_stop_at_last_layers": OptionInfo(1, "Clip skip", gr.Slider, {"minimum": 1, "maximum": 12, "step": 1}, infotext="Clip skip").link("wiki", "https://github.com/AUTOMATIC1111/stable-diffusion-webui/wiki/Features#clip-skip").info("ignore last layers of CLIP network; 1 ignores none, 2 ignores one layer"),
     "upcast_attn": OptionInfo(False, "Upcast cross attention layer to float32"),
     "randn_source": OptionInfo("GPU", "Random number generator source.", gr.Radio, {"choices": ["GPU", "CPU", "NV"]}, infotext="RNG").info("changes seeds drastically; use CPU to produce the same picture across different videocard vendors; use NV to produce same picture as on NVidia videocards"),