video generation result on lingbot-va-posttrain-robotwin

[21Broccoli]

Thanks for the great work!

I evaluated the tasks in RoboTwin using the lingbot-va-posttrain-robotwin checkpoint, and the success rate is impressively high!

However, while checking the visualizations, I noticed that the backgrounds of the generated prediction videos always seem to have a green tint. For example, please see the open_microwave and rank_rgb videos attached below.

Is this normal behavior? For context, I am calling the decode_one_video function in wan_va/wan_va_server.py.

0_Engage_the_microwave_using_the_left_arm_to_open._True.mp4

pred.mp4

[lilin-hitcrt]

Thanks for your feedback! The green tint you're seeing is not normal behavior.

Could you first try running the I2AV example to see if the issue persists?

Also, could you provide more details about your decode process? Normally, after running the code, latent files (latents_*.pt) are saved in output_dir/rel/. The standard process is to load these files, combine the latents together, and then decode them with the VAE. Are you following this process, or are you doing something different?

[21Broccoli]

Thank you so much! Through the mentioned process I can generate the prediction video in normal behavior!

pred.mp4