Stable Diffusion Black Images and NaN Error

How to fix it

Try these in order — the most common fix is first.

1. 1. Force full-precision VAE

   Copy steps
   1. Edit webui-user.bat
   2. Set COMMANDLINE_ARGS=--no-half-vae
   3. Restart the web UI
   Did this fix it? Yes No

2. 2. If NaNs come from the Unet

   Copy steps
   1. Add --no-half as well (uses more VRAM)
   2. Or try --upcast-sampling first on cards that support it
   Did this fix it? Yes No

3. 3. SDXL specific

   Copy steps
   1. Use the fp16-fixed SDXL VAE (sdxl-vae-fp16-fix) in Settings > Stable Diffusion > SD VAE
   2. Avoid --disable-nan-check: it only turns off the warning
   Did this fix it? Yes No

Symptoms

Output is a solid black or green image, or the UI shows 'NansException: A tensor with all NaNs was produced in Unet / VAE'.

Why it happens

• fp16 precision overflow, especially the original SDXL VAE
• GPU that does not support half precision well
• Incompatible VAE or checkpoint
• Using --disable-nan-check, which hides the error but leaves black images

This guide also covers

• stable diffusion black images
• a tensor with all nans was produced
• stable diffusion green image

The same problem in other apps

All ai & editing tools not working guides →

• Leonardo.ai Tokens Ran Out: Why and What to Do (Leonardo.ai)

Other Stable Diffusion errors

• Stable Diffusion CUDA Out of Memory Error
• Stable Diffusion "Torch is not able to use GPU" Error
• Stable Diffusion Model Failed to Load, Exiting
• Stable Diffusion Very Slow on Mac (M1, M2, M3)
• Stable Diffusion Not Using AMD GPU (Slow, CPU Only)
• Stable Diffusion xformers Error After Update
• Stable Diffusion "Couldn't install torch" Error
• Stability AI API 402 Insufficient Credits Error