Fix mask handling for batch generation in HunyuanVideo example

[tjkemp]

Summary

Batch inference was using the mask from first item (encoder_attention_mask[0]) in hunyan_video_usp_example.py, which works for bs=1 but cuts off masks for longers prompts and produces artifacts. This PR provides a fix.

How to reproduce

1. Apply patch to make script save all videos

diff --git a/examples/hunyuan_video_usp_example.py b/examples/hunyuan_video_usp_example.py
index 03856f1..a9041ac 100644
--- a/examples/hunyuan_video_usp_example.py
+++ b/examples/hunyuan_video_usp_example.py
@@ -297,7 +291,7 @@ def main():
         guidance_scale=input_config.guidance_scale,
         generator=torch.Generator(device="cuda").manual_seed(
             input_config.seed),
-    ).frames[0]
+    )
 
     end_time = time.time()
     elapsed_time = end_time - start_time
@@ -311,9 +305,10 @@ def main():
     )
     if is_dp_last_group():
         resolution = f"{input_config.width}x{input_config.height}"
-        output_filename = f"results/hunyuan_video_{parallel_info}_{resolution}.mp4"
-        export_to_video(output, output_filename, fps=15)
-        print(f"output saved to {output_filename}")
+        for idx, frames in enumerate(output.frames, start=1):
+            output_filename = f"results/hunyuan_video_{idx:02d}_{parallel_info}_{resolution}.mp4"
+            export_to_video(frames, output_filename, fps=15)
+            print(f"output saved to {output_filename}")
 
     if get_world_group().rank == get_world_group().world_size - 1:
         print(

2. Run an example with an added second prompt.

mkdir -p results && torchrun --nproc_per_node=2 examples/hunyuan_video_usp_example.py --model tencent/HunyuanVideo --ulysses_degree 2 --num_inference_steps 30 --warmup_steps 0 --prompt "A husky puppy plays with its own tail." "Two Siamese cats eat sushi from a plate." --height 320 --width 512 --num_frames 61 --enable_tiling --enable_model_cpu_offload

3. Compare results

The second video before this fix:

The second video after applying this fix:

Notes

Unrelated to this change, I’m seeing:

[rank1]:   File "/app/xDiT/examples/hunyuan_video_usp_example.py", line 297, in main
[rank1]:     guidance_scale=input_config.guidance_scale,
[rank1]:                    ^^^^^^^^^^^^^^^^^^^^^^^^^^^
[rank1]: AttributeError: 'InputConfig' object has no attribute 'guidance_scale'

[feifeibear]

guidance_scale error probably comes from capability of diffusers version

[tjkemp]

Is there anything else I could adjust to help get this PR ready for merge?