sculmh

你可以尝试将训练的分辨率调低、trainable_modules只微调一部分模块而不是'all'、cache_latents设为True等方法降低显存

我之前试过在低分辨率上微调，高分辨率上推理，没有发现效果有明显下降。不过没有进行规模化的实验验证。

不是lora。SVD这部分工作我参与不多，不是很熟。

你用的图片还是视频微调motion？

Demos are the result from animate_anything rather than animate_anything_SVD.

In consideration of efficiency, if we opt to employ motion representation in the RGB format, it will be required to transform `predict_x0` back into the RGB space in the line...

Please refer to VideoJsonDataset.__get_item__ (util/dataset.py), Videojson contains list of dict like {"caption": "xxx", "video": "path"}. [{"caption": "Motion around stack of large mesh bags in food warehouse ', 'video': '000051_000100/1066665592.mp4"}]