[bug] Solution may affect accuracy

[LukeLIN-web]

seems in many cases, like chess, first image height is too low, it may affect accuracy. Not sure this is from task side or model crop it ?
We need to resize image for each model.!!!!!!!

• add resize/crop code for each model 。 first crop to recommend height/width ratio, then resize to recommend resolution.

[LukeLIN-web]

since has "absolute position embedding
RoPE / 2D-RoPE / 3D-RoPE
learned spatial grid embedding " height may have strong impact on video model performance. otherwise rope cannot Extrapolate to image solution.

e.g. wan2.2 would not crop, it resize the input image. https://github.com/Wan-Video/Wan2.2/blob/e9783574/wan/image2video.py

• crop may cut the key point in our reasoning task. How to solve this?

[LukeLIN-web]

Wan2.2 is 1280 x 720
sora2 1280 x 720 ， sora2 pro 1792x1024.

let us generate 1792x1024 now , crop it to each input.