The orig_wan implementation has a different block mask

[m-richa]

I used the same image in both the main branch and the Benchmark branch. The block mask had a different shape in both, smaller in Benchmark. There was also a quality degradation in the Benchmark repository.

[haochengxi]

Can you explain this in detail, or take a screenshot for the attention mask in the main branch and the Benchmark branch?
We might slightly change the behavior of the attention mask when doing ablation and forgot to change them back.

[m-richa]

Ok, got it. The benchmark code does not reshape_and_crop the image into 1280720 / 7201280 and directly uses the original image shape so I was getting different masks. But I am still seeing a quality degradation between the Benchmark and the main branch.

I thought it was because some params are changed block_thresh, first frame_sink, or using flex instead of SDPA for full masks. But I can't recover the original quality. It is mostly evident in the last few frames.

Do you have any idea what could be the cause?

[haochengxi]

We will come back to this problem later to see what caused this issue. Are you using the wan_i2v_inference.py or orig_wan_generate.py?