update

app.py

@@ -105,7 +105,9 @@ with gr.Blocks(css=css) as demo:
         gr.Markdown(
             f"""# MV-Adapter [Text-to-Multi-View]
 Generate 768x768 multi-view images using {base_model} <br>
-[[page](https://huanngzh.github.io/MV-Adapter-Page/)] [[repo](https://github.com/huanngzh/MV-Adapter)] [Tips: please follow the prompt template in [usage-guidelines](https://huggingface.co/cagliostrolab/animagine-xl-3.1#usage-guidelines)]
+[[page](https://huanngzh.github.io/MV-Adapter-Page/)] [[repo](https://github.com/huanngzh/MV-Adapter)] <br>
+Also check our other demos: [Text-to-Multiview (General)](https://huggingface.co/spaces/VAST-AI/MV-Adapter-T2MV-SDXL) | [Text-to-Multiview (Anime)](https://huggingface.co/spaces/huanngzh/MV-Adapter-T2MV-Anime) | [Image-to-Multiview](https://huggingface.co/spaces/VAST-AI/MV-Adapter-I2MV-SDXL) <br>
+**Tips:** please follow the prompt template in [usage-guidelines](https://huggingface.co/cagliostrolab/animagine-xl-3.1#usage-guidelines)
         """
         )
 

mvadapter/pipelines/pipeline_mvadapter_i2mv_sdxl.py

@@ -906,6 +906,14 @@ class MVAdapterI2MVSDXLPipeline(StableDiffusionXLPipeline, CustomAdapterMixin):
                 use_mv=True,
                 use_ref=True,
             ),
+            set_cross_attn_proc_func=lambda name, hs, cad, ap: self_attn_processor(
+                query_dim=hs,
+                inner_dim=hs,
+                num_views=num_views,
+                name=name,
+                use_mv=False,
+                use_ref=False,
+            ),
         )
 
         # copy decoupled attention weights from original unet

mvadapter/pipelines/pipeline_mvadapter_t2mv_sdxl.py

@@ -747,6 +747,14 @@ class MVAdapterT2MVSDXLPipeline(StableDiffusionXLPipeline, CustomAdapterMixin):
                 use_mv=True,
                 use_ref=False,
             ),
+            set_cross_attn_proc_func=lambda name, hs, cad, ap: self_attn_processor(
+                query_dim=hs,
+                inner_dim=hs,
+                num_views=num_views,
+                name=name,
+                use_mv=False,
+                use_ref=False,
+            ),
         )
 
         # copy decoupled attention weights from original unet