search

cswry / OSEDiff

[NeurlPS2024] One-Step Effective Diffusion Network for Real-World Image Super-Resolution
Apache License 2.0
215 stars 12 forks source link

The results are "unfairly" much more natural and detailed without the tiling function #32

Closed zelenooki87 closed 2 months ago

zelenooki87 commented 2 months ago

However, even when I implement fp16 and remove the tiling parts of the code to make it work without splitting the images into tiles, I still get an out of memory error on anything above 800x600 resolution. Why does tiling degrade the results so much and make them appear somewhat "artificial"?

AIisCool commented 2 months ago

@zelenooki87 Can you share the revised code without the tiling? I am curious how different it looks compared to a titled image. Thanks!

zelenooki87 commented 2 months ago

After further testing, I think your best option is to experiment only with these parameters: --latent_tiled_size --latent_tiled_overlap

Bigger size-better fidelity ( more looking to original) and lower maybe sometimes too "artificial"

Fully disabled tiling is not good idea as I mean before. Sorry

prashantg445 commented 1 week ago

Hey @zelenooki87, Thanks for your insights, indeed they work. Just wanted to add "wavelet" color fix also help in fixing this issue.

zelenooki87 commented 1 week ago

@prashantg445 Try S3Diff project, it is much better in terms of Fidelity. https://github.com/ArcticHare105/S3Diff/