magic-research / magic-edit Goto Github PK

Have you tried other experiments such as not disentangle algorithms to provide some straightforward comparisons like image?
As I think, the best point of this method is less training than other methods and is quite like AnimateDiff and stablediffusion-controlnet-pose and Personalization methods (which I believe lots of people have already used the combination of these as mentioned in your paper) but this work may not be so much convincing to say dissentanglement is the key to high-fidelity temporally coherent.

Thanks for your great work!
May I ask how long video length this method can support？

Hi,
In the technical report you mentioned you use DDIM inversion for video out painting. There are a few different papers to do DDIM inversion. Which way was tried?
Was this used - https://huggingface.co/docs/diffusers/v0.14.0/en/api/schedulers/ddim_inverse ?

有具体的发布时间吗，注意到，大家对此都充满了期待。
Is there a specific release time? I noticed that everyone is looking forward to it.
具体的な発売時期はありますか?みんな楽しみにしています。
구체적인 출시 시간이 있습니까? 모두가 기대하고 있다는 것을 알았습니다.
มีเวลาวางจำหน่ายที่แน่นอนหรือไม่ โปรดทราบว่าทุกคนตั้งตารอคอยสิ่งนี้
Есть конкретное время выпуска? Я заметила, что все с нетерпением ждут этого.
Có thời gian phát hành cụ thể không, lưu ý rằng mọi người đều đầy kỳ vọng vào điều này.

As I understand you making following steps:

1. Fine tune dreambooth on all frames
2. Fine tune animatediff motion module
3. Extract controlnet from video
4. Combine them to infer

Can you share hyperparameters of finetuning?

1. Do you finetune LoRA or Dreambooth? How many steps, what lr?
2. Do you finetune all module or selected layers?

great work!
can you provide ablation results of w/o Stage II-A, w/o Stage II-B & w/o Stage II-A&B? so that i can have an intuitive feeling about how these two modules contribute to temporal coherent respectively.
thx

when release code?