test v2

Author: hlky

src/diffusers/loaders/transformer_flux.py

@@ -64,8 +64,10 @@ def _convert_ip_adapter_image_proj_to_diffusers(self, state_dict, low_cpu_mem_us
             # IP-Adapter
             # TODO: fix for XLabs-AI/flux-ip-adapter-v2
             num_image_text_embeds = 4
+            if state_dict["proj.weight"].shape[0] == 65536:
+                num_image_text_embeds = 16
             clip_embeddings_dim = state_dict["proj.weight"].shape[-1]
-            cross_attention_dim = state_dict["proj.weight"].shape[0] // 4
+            cross_attention_dim = state_dict["proj.weight"].shape[0] // num_image_text_embeds
 
             with init_context():
                 image_projection = ImageProjection(
@@ -124,9 +126,11 @@ def _convert_ip_adapter_attn_to_diffusers(self, state_dicts, low_cpu_mem_usage=F
                 num_image_text_embeds = []
                 for state_dict in state_dicts:
                     if "proj.weight" in state_dict["image_proj"]:
+                        num_image_text_embed = 4
+                        if state_dict["image_proj"]["proj.weight"].shape[0] == 65536:
+                            num_image_text_embed = 16
                         # IP-Adapter
-                        # TODO: change for XLabs-AI/flux-ip-adapter-v2
-                        num_image_text_embeds += [4]
+                        num_image_text_embeds += [num_image_text_embed]
 
                 with init_context():
                     attn_procs[name] = attn_processor_class(

src/diffusers/pipelines/flux/pipeline_flux.py

@@ -449,6 +449,8 @@ def encode_image(self, image, device, num_images_per_prompt):
 
         image = image.to(device=device, dtype=dtype)
         image_embeds = self.image_encoder(image).image_embeds
+        # flux-ip-adapter-v2
+        image_embeds = torch.mean(image_embeds)
         image_embeds = image_embeds.repeat_interleave(num_images_per_prompt, dim=0)
         return image_embeds