Batching/Sliding Window about flatten

[yrcong]

Hello logtd,

I am glad that you have interest in our model and welcome to integrate FLATTEN into ComfyUI!

For processing a video in chunks, TBH, we didn't test it. inter_frame and get_slide_window_indices are used for something else and left.

I would like to suggest that you could do in this way:
Let's say there is a 100-frame video, you could chunk it into three clips: [1-40], [30-70], [60-100] and process them with FLATTEN in batch. Since there are overlaps between these chunks, the visual appearance should be somehow consistent.

We will move on to optimize FLATTEN to make it more efficient:D

best regards,
Yuren

[logtd]

Thanks for your reply!

I've implemented a naive batching solution and it works somewhat. Will try to improve it over time.

One interesting thing I've found about Flatten is that while sampling you can use a sampling algorithm that adds additional noise as long as the noise is guided by optical flow/trajectories through frames. It allows the model to produce videos that can deviate from the source input but maintain consistency. Here's a simple (yet slow) implementation of it https://github.com/logtd/ComfyUI-FLATTEN/blob/main/utils/flow_noise.py

Again, thanks for open sourcing this code. Going to close the issue but feel free to reopen.

[yrcong]

Thanks for the information! Please let me know if you have other concerns. After ECCV I will have more time;D