Where is the code related to “variable resolutions and aspect ratios” ?

[EmmaThompson123]

I'm very interested in this feature, could you please tell me where the relevant code is ?

[sculmh]

it is not implemented, but you can refer to https://github.com/hpcaitech/Open-Sora/blob/main/opensora/datasets/bucket.py, which buckets the training videos based on (resolution, num_frame, aspect_ratio), using data from the same bucket in one batch during training.

[zhenghao977]

@EmmaThompson123 You can refer to our latest work Tora: https://github.com/alibaba/Tora