Initial support for finetuning qwen image. Will only work with caching for now, need to add controls everywhere.

toolkit/stable_diffusion_model.py

@@ -217,6 +217,9 @@ class StableDiffusion:
         # a list of trainable multistage boundaries
         self.trainable_multistage_boundaries: List[int] = [0]
         
+        # set true for models that encode control image into text embeddings
+        self.encode_control_in_text_embeddings = False
+        
     # properties for old arch for backwards compatibility
     @property
     def is_xl(self):
@@ -2356,6 +2359,7 @@ class StableDiffusion:
             long_prompts=False,
             max_length=None,
             dropout_prob=0.0,
+            control_images=None,
     ) -> PromptEmbeds:
         # sd1.5 embeddings are (bs, 77, 768)
         prompt = prompt