-
Hello, |
Beta Was this translation helpful? Give feedback.
Replies: 3 comments
-
https://github.com/pkuliyi2015/multidiffusion-upscaler-for-automatic1111 |
Beta Was this translation helpful? Give feedback.
-
Thanks, I know MultiDiffusion and use it. It's not exactly a substitute to solve Hires Fix running out of CUDA memory, but ok. I can't get MultiDiffusion installed under SD.Next however. |
Beta Was this translation helpful? Give feedback.
-
Diffusers Settings -> VAE Tiling Tho i would also recommend HyperTile UNet to speed things up. Never OOM is pretty much the same thing as Sequential Offload + VAE Tiling. Generating a 6552x6552 image with VAE Tiling + Dynamic Attention SDP + Sequential Offload takes 5-6 GB with 7.5 GB peaks: |
Beta Was this translation helpful? Give feedback.
Diffusers Settings -> VAE Tiling
Compute Settings -> Attention optimization method -> Dynamic Attention SDP
Tho i would also recommend HyperTile UNet to speed things up.
Never OOM is pretty much the same thing as Sequential Offload + VAE Tiling.
Sequential Offload has pretty high performance hit.
You can enable Sequential Offload either from Diffusers Settings or using --lowvram.
Generating a 6552x6552 image with VAE Tiling + Dynamic Attention SDP + Sequential Offload takes 5-6 GB with 7.5 GB peaks: