Home Artists Posts Import Register

Content

Here we go, after so many updates about it.

There is the new two models, hope they are good. For now they are a little slower from the default models, since I already had a hard time making they somewhat decent to use. In the future I can try to optimize them more.


  • Fast Interpolation. This is a improved version of "Alternated Interpolation". For RL it usually not great, but it can work well on some animations. The advantage is that you can interpolate in 15X the same speed of 3X (It work with uneven numbers)
  • New Model Anime and Real Life are the new models.
  • Fixed a bunch of options
  • Had to remove some options until to fix it, like alpha channel. I could train a model with alpha, but the dataset would be a pain to make.
  • I am not sure how well the new models will work with 10bit videos, but I learned a bunch of stuff now that can improve it for 10bit in the next versions.
  • Let me know if you think the new models are good or crap o/

Files

RIFE-App 2.5.rar

Comments

Cheshire Cad

Unfortunately, the "Create a Slow Motion output" isn't applying to any GIFs being made. It only applies to the MP4 videos.

DAINAPP

Yeah, that is current a limitation. I need to improve the output for some combinations.

Anonymous

RuntimeError: CUDA out of memory. Tried to allocate 12.74 GiB (GPU 0; 24.00 GiB total capacity; 607.85 MiB already allocated; 20.47 GiB free; 1.31 GiB reserved in total by PyTorch)

Anonymous

I met the error above when I run the RIFE. Also, I have the same error when I train the cyclegan open source project. Want to know how to fix it, thank you.

Snake Plissken

Does using the player reduce processing speed? Same with using YUV? I’m guessing YUV will give more accuracy?

Snake Plissken

Forgot if I mentioned this but seems the concat and YUV options dont play nice, ends up making a 4 panel video

Snake Plissken

is it possible to have either a 90 degree rotate option OR preserve the rotation flag in the meta data? For video recorded in profile from a phone. Talking about this metadata https://blog.addpipe.com/mp4-rotation-metadata-in-mobile-video-files/ - best solution I think is to check if video has this flag, and then add it to the final output(s) (if concat is flagged then it needs to be done differently)

DAINAPP

Actually I am remaking the YUV system, it was trash, it should be better on my next update. For Hi-Res, player should slowdown the app, yeah. I may improve the memory usage in the future.

DAINAPP

This used to be a problem, it look like it's back, dang it. Thanks for letting me know.

DAINAPP

That is a good point, never ocurred to me to check for the rotation flag. I will put it in my TODO list, but it may take a while, since it's not something super important.

Anonymous

For my use case, I'm finding the RL model is far more accurate than previous models and is doing an excellent job of interpolating frames with fewer artefacts than before. Excellent work!

DAINAPP

Thank you, yours is the first review someone gave to my model, so thank you very much. The new animation model is already working great compared to the release model, so it's very possible that the new RL model will work better as well. I also improved the code for YUV and color correction output.