Experimental clip skip interpolation

modules/prompt_parser_diffusers.py

@@ -1,4 +1,5 @@
 import os
+import math
 import time
 import typing
 import torch
@@ -17,15 +18,20 @@ def compel_hijack(self, token_ids: torch.Tensor,
                                             attention_mask,
                                             output_hidden_states=needs_hidden_states,
                                             return_dict=True)
+    if not needs_hidden_states:
+        return text_encoder_output.last_hidden_state
     normalized = self.returned_embeddings_type > 0
-    if normalized and needs_hidden_states:
-        clip_skip_hidden_state = text_encoder_output.hidden_states[-abs(self.returned_embeddings_type)]
-        return self.text_encoder.text_model.final_layer_norm(clip_skip_hidden_state)
-    if needs_hidden_states:
-        clip_skip_hidden_state = text_encoder_output.hidden_states[-abs(self.returned_embeddings_type)]
-        return clip_skip_hidden_state
+    clip_skip = math.floor(abs(self.returned_embeddings_type))
+    interpolation = abs(self.returned_embeddings_type) - clip_skip
+    if interpolation:
+        hidden_state = (1 - interpolation) * text_encoder_output.hidden_states[-clip_skip] + interpolation * text_encoder_output.hidden_states[-(clip_skip+1)]
+    else:
+        hidden_state = text_encoder_output.hidden_states[-clip_skip]
+    if normalized:
+        hidden_state = self.text_encoder.text_model.final_layer_norm(hidden_state)
+    return hidden_state
+
 
-    return text_encoder_output.last_hidden_state
 
 
 EmbeddingsProvider._encode_token_ids_to_embeddings = compel_hijack

modules/ui_sections.py

@@ -109,7 +109,7 @@ def create_advanced_inputs(tab):
                 diffusers_guidance_rescale = gr.Slider(minimum=0.0, maximum=1.0, step=0.05, label='Guidance rescale', value=0.7, elem_id=f"{tab}_image_cfg_rescale", visible=shared.backend == shared.Backend.DIFFUSERS)
                 diffusers_sag_scale = gr.Slider(minimum=0.0, maximum=1.0, step=0.05, label='Attention guidance', value=0.0, elem_id=f"{tab}_image_sag_scale", visible=shared.backend == shared.Backend.DIFFUSERS)
             with gr.Row():
-                clip_skip = gr.Slider(label='CLIP skip', value=1, minimum=0, maximum=12, step=1, elem_id=f"{tab}_clip_skip", interactive=True)
+                clip_skip = gr.Slider(label='CLIP skip', value=1, minimum=0, maximum=12, step=0.1, elem_id=f"{tab}_clip_skip", interactive=True)
         with gr.Group():
             gr.HTML('<br>')
             with gr.Row():

scripts/xyz_grid.py

@@ -233,7 +233,7 @@ axis_options = [
     AxisOption("CFG End", float, apply_field("cfg_end")),
     AxisOption("Variation seed", int, apply_field("subseed")),
     AxisOption("Variation strength", float, apply_field("subseed_strength")),
-    AxisOption("Clip skip", int, apply_clip_skip),
+    AxisOption("Clip skip", float, apply_clip_skip),
     AxisOption("Denoising strength", float, apply_field("denoising_strength")),
     AxisOption("Prompt order", str_permutations, apply_order, fmt=format_value_join_list),
     AxisOption("Model dictionary", str, apply_dict, fmt=format_value, cost=1.0, choices=lambda: ['None'] + list(sd_models.checkpoints_list)),