fix vis

.gitignore

@@ -42,4 +42,5 @@ hw*
 257f*
 513f*
 taming*
-221hw*
+221hw*
+65x512x512

opensora/models/diffusion/latte/modeling_latte.py

@@ -401,7 +401,7 @@ class LatteT2V(ModelMixin, ConfigMixin):
                 encoder_hidden_states = torch.cat([encoder_hidden_states_video, encoder_hidden_states_image], dim=1)
                 encoder_hidden_states_spatial = rearrange(encoder_hidden_states, 'b f t d -> (b f) t d').contiguous()
             else:
-                encoder_hidden_states_spatial = repeat(encoder_hidden_states, 'b t d -> (b f) t d', f=frame).contiguous()
+                encoder_hidden_states_spatial = repeat(encoder_hidden_states, 'b 1 t d -> (b f) t d', f=frame).contiguous()
 
         # prepare timesteps for spatial and temporal block
         timestep_spatial = repeat(timestep, 'b d -> (b f) d', f=frame + use_image_num).contiguous()

opensora/sample/pipeline_videogen.py

@@ -705,6 +705,10 @@ class VideoGenPipeline(DiffusionPipeline):
                 # broadcast to batch dimension in a way that's compatible with ONNX/Core ML
                 current_timestep = current_timestep.expand(latent_model_input.shape[0])
 
+                if prompt_embeds.ndim == 3:
+                    prompt_embeds = prompt_embeds.unsqueeze(1)  # b l d -> b 1 l d
+                # if prompt_attention_mask.ndim == 2:
+                #     prompt_attention_mask = prompt_attention_mask.unsqueeze(1)  # b l -> b 1 l
                 # predict noise model_output
                 noise_pred = self.transformer(
                     latent_model_input,