search

csslc / CCSR

Official codes of CCSR: Improving the Stability of Diffusion Models for Content Consistent Super-Resolution
https://csslc.github.io/project-CCSR/
390 stars 30 forks source link

Nothing in outputs for input 640x480, after long time... 24GB VRAM #1

Open zelenooki87 opened 6 months ago

zelenooki87 commented 6 months ago

Hi. Thank you for wonderful project. I was able to evaluate model on Windows. but only for images <320x240

I have RTX 3090 24GB VRam gpu. Is it expected to be such a slow and without results (Real world model)? Do you recommend try code in Ubuntu? thank you

csslc commented 6 months ago

Hi, thanks for your interest in our work! In our inference code of 'inference_ccsr.py', the input is resized to 512 by default. Under the default settings, the size of the input image should not be affected. Could you provide more information, such as the error message when running the code?

zelenooki87 commented 6 months ago

That was because of Out of memory. Can we use your model in tile mode? please advice

csslc commented 6 months ago

You can use the tiling function by modifying the 'tiled' from 'store_true' to 'store_false' in 'inference_ccsr.py'. You can also choose appropriate 'tile_size' and 'tile_stride' according to your GPU memory and running time. However, more border artifacts may be introduced in this function.