Update pipeline_stable_diffusion_3_ipa.py
Browse files
pipeline_stable_diffusion_3_ipa.py
CHANGED
@@ -1175,15 +1175,7 @@ class StableDiffusion3Pipeline(DiffusionPipeline, SD3LoraLoaderMixin, FromSingle
|
|
1175 |
image_prompt_embeds_list.append(image_prompt_embeds_5)
|
1176 |
|
1177 |
# Concatenate the image embeddings
|
1178 |
-
|
1179 |
-
embedding_dim = concatenated_embeds.shape[-1] # Get the embedding dimension
|
1180 |
-
|
1181 |
-
total_embedding_dim = concatenated_embeds.shape
|
1182 |
-
linear_layer = nn.Linear(embedding_dim * len(image_prompt_embeds_list), embedding_dim * len(image_prompt_embeds_list), dtype=self.dtype).to(self.device)
|
1183 |
-
clip_image_embeds = linear_layer(concatenated_embeds)
|
1184 |
-
|
1185 |
-
# Add a ReLU activation for non-linearity (optional)
|
1186 |
-
#combined_embeds = torch.relu(combined_embeds)
|
1187 |
|
1188 |
# 4. Prepare timesteps
|
1189 |
timesteps, num_inference_steps = retrieve_timesteps(self.scheduler, num_inference_steps, device, timesteps)
|
|
|
1175 |
image_prompt_embeds_list.append(image_prompt_embeds_5)
|
1176 |
|
1177 |
# Concatenate the image embeddings
|
1178 |
+
clip_image_embeds = torch.mean(torch.stack(image_prompt_embeds_list), dim=0)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1179 |
|
1180 |
# 4. Prepare timesteps
|
1181 |
timesteps, num_inference_steps = retrieve_timesteps(self.scheduler, num_inference_steps, device, timesteps)
|