How to truly avoid NaN errors

[freecoderwaifu]

1. Don't use the NAI VAE, it's broken, it's time to move to better VAEs. The NAI VAE is the root cause for the majority of NaN errors. It doesn't help that the VAE is distributed with a myriad of anime merges and just renamed. Anything VAE? Still the NAI VAE, Abyss Orange Mix (AOM) VAE? Still the NAI VAE. It also doesn't help that some checkpoints bake in the NAI VAE.

Use the WD 1.4 VAE for anime and 2D (you can convert that ckpt to safetensors with the Toolkit extension or just the regular merger, or just use it as is)
https://huggingface.co/hakurei/waifu-diffusion-v1-4/tree/main/vae

Use the SD 1.5 VAE for realism, 3D CGI and less 2D models. (get the safetensors)
https://huggingface.co/stabilityai/sd-vae-ft-mse-original/tree/main

Try other VAEs.
https://civitai.com/search/models?sortBy=models_v3&query=vae

If any of the VAEs from Civitai gives you a NaN error, it's just the NAI VAE renamed.

If your checkpoint has the NAI VAE baked in, manually set a different VAE in Settings, instead of setting it to Automatic or None.

If you absolutely must use the NAI VAE for whatever reason, add --no-half-vae to your webui-user.bat args.

https://github.com/AUTOMATIC1111/stable-diffusion-webui/wiki/Command-Line-Arguments-and-Settings

For SDXL, try the SDXL fixed VAE instead of the official one
https://huggingface.co/madebyollin/sdxl-vae-fp16-fix/tree/main

2. Try not to mix FP16 with FP32, the webUI by default converts checkpoints to FP16 when loading them, and sometimes mixing precision with other submodels that use FP32 could also trigger a NaN.

If using ControlNet, try the FP16 converted models instead.
https://huggingface.co/comfyanonymous/ControlNet-v1-1_fp16_safetensors/tree/main

If using Adetailer, I'm not sure if their models are FP32 or FP16, but there could FP16 alternatives.

Try --upcast-sampling or --no-half. Note that it will be slower, since it will force full precision instead of half.

https://github.com/AUTOMATIC1111/stable-diffusion-webui/wiki/Command-Line-Arguments-and-Settings

3. Your graphics card might not fully support FP16 or mixing precision. This applies to some older GTX cards. For AMD cards, even if they support FP16 and mixed precision, it could also be an issue with them due to their Windows implementation. Same with CPU only mode due to its particularities.

4. Adding --disable-nan-check only disables the warning, not the root cause.

[rigaorenke]

niubi