.ai

the best open source ai models

Guides

Rankings

FAQ

Upload

Automatic1111 (Stable Diffusion WebUI) errors and fixes

Q: Can A1111 run Flux?

Not well. Flux support lives in Forge and ComfyUI. The Flux models on this site name the UIs they work with.

Q: Torch is not able to use GPU

CPU-only torch, old driver, or non-NVIDIA card. Update the driver, delete venv and relaunch to reinstall torch with CUDA. AMD on Windows: use the DirectML fork or ZLUDA; on Linux use ROCm.

Q: NansException with SDXL

fp16 VAE overflow. Download sdxl-vae-fp16-fix, select it in SD VAE, or add --no-half-vae.

Q: CUDA out of memory with hires fix

The upscaled pass needs several times the memory. Lower the upscale factor, use a latent upscaler, switch Tiled VAE on, or run the hires pass as img2img afterwards.

Q: Couldn't install torch

Python 3.11 or 3.12 instead of 3.10. Install Python 3.10.6, point webui-user.bat at it with set PYTHON=path\python.exe, delete venv, relaunch.

Q: Stable diffusion model failed to load

Half-downloaded or renamed file. Re-download, check size against the model page, keep the original extension.

Topic: Errors & fixes

By Captain

Published 2026-05-14

Overview

AUTOMATIC1111's Stable Diffusion WebUI is still the most installed interface for

SD 1.5

and

SDXL

, and its error messages are well known by now. Most of them come from four causes: the Python or torch install, a card that is too small for the settings, a model file that does not match the settings (VAE, half precision, SDXL versus 1.5), and extensions that broke after an update.

The WebUI reads its start-up options from webui-user.bat (Windows) or webui-user.sh (Linux and Mac) in the COMMANDLINE_ARGS line. Most fixes below are a flag added to that line. After a change, close the console window and start the WebUI again; a browser refresh is not enough.

If the WebUI is the bottleneck rather than the model, the newer

Forge

fork is a drop-in replacement with much lower memory use (its own guide is linked below), and the same SD 1.5 and SDXL checkpoints run on the cloud through

PirateDiffusion

's /render command or

BitVector

's web app without any install.

Reference

Name	Type	What it is
Torch is not able to use GPU	error	torch was installed without CUDA, the driver is too old, or the card is AMD/Intel. Reinstall torch with CUDA or add the right flags for your hardware.
NansException: A tensor with all NaNs was produced	error	Half precision overflow in the VAE or the Unet, typical with SDXL and with some merged checkpoints. Use --no-half-vae, then --no-half, or the fp16-fix VAE.
RuntimeError: CUDA out of memory	error	Resolution, batch, hires fix or ControlNet exceed VRAM . Add --medvram (or --medvram-sdxl) and lower the hires upscale.
Couldn't install torch / No matching distribution	error	The wrong Python version (the WebUI wants 3.10.x) or a blocked network. Install Python 3.10.6, add it to PATH, delete the venv folder and start again.
Stable diffusion model failed to load	error	A corrupt or incomplete checkpoint , a LoRA in the checkpoint folder, or a config mismatch. Re-download, and keep LoRAs in models/Lora.
AttributeError: 'NoneType' object has no attribute ...	error	An extension that no longer matches the WebUI version. Disable extensions, update them one by one.
xformers: WARNING ... / version mismatch	warning	xformers compiled against another torch. Remove --xformers and use --opt-sdp-attention, or reinstall xformers with --reinstall-xformers once.
DefaultCPUAllocator: not enough memory	error	System RAM ran out while loading a 6 GB SDXL checkpoint. Close other apps, enlarge the page file, or convert the checkpoint to fp16.
Connection errored out / Gradio timeout	symptom	The Python process crashed or stalled. Read the console; a crash during VAE decode is memory.
Images are black or green	symptom	Half precision VAE on cards that cannot do it (GTX 16xx, some 10xx). Add --no-half-vae or --precision full --no-half.

No install required - run AI models on the cloud

BitVector Prism is the easiest way to start: pick a model, type a prompt and generate in a clean web app, with nothing to configure. BitVector is also available on Discord and on the web (SpyGlass).

Try BitVector

Step by step

Confirm Python 3.10.x and a recent NVIDIA driver (the console prints the torch and CUDA versions at start-up; CUDA should not be "cpu").
Edit webui-user.bat and set COMMANDLINE_ARGS. Start with: --xformers or --opt-sdp-attention, plus --medvram for 8 GB cards, --lowvram for 4 to 6 GB, --no-half-vae for SDXL.
Put files where the WebUI expects them: checkpoints in models/Stable-diffusion, LoRAs in models/Lora, embeddings in the embeddings folder, VAEs in models/VAE, ControlNet models in extensions/sd-webui-controlnet/models or models/ControlNet.
For a NansException, first add --no-half-vae; if it persists add --no-half; with SDXL use the sdxl-vae-fp16-fix VAE instead of the baked one. Do not tick "Upcast cross attention" unless the message asks for it.
For a memory error, reduce before you upgrade: batch size 1, no hires fix above 1.5x, 512x768 for SD 1.5 and 1024x1024 for SDXL, Tiled VAE from the multidiffusion extension for big outputs.
After an update that broke things, start with --disable-all-extensions. If it runs, re-enable in Settings > Extensions in halves, Apply and restart UI each time, until you find the offender; update or remove it.
If the install is damaged, delete the venv folder (only that folder) and start webui-user.bat again; it rebuilds the environment in a few minutes and keeps your models and outputs.
When a card is simply too small for SDXL or ControlNet, use the same checkpoint hosted: PirateDiffusion renders SD 1.5 and SDXL models with /render and the model's trigger word from this site, and BitVector runs them in the browser.

Examples

webui-user.bat for an 8 GB card with SDXL

set COMMANDLINE_ARGS=--opt-sdp-attention --medvram-sdxl --no-half-vae --api

GTX 1660 / 1650 (black images)

set COMMANDLINE_ARGS=--lowvram --precision full --no-half --opt-sdp-attention

Rebuild the environment

rmdir /s /q venv

webui-user.bat

Tips

A1111 is maintained slowly; Forge (its fork) has most fixes built in and uses half the memory. The Forge guide on this site covers the differences.
Keep the Settings > Stable Diffusion > SD VAE on Automatic and give each checkpoint its own VAE by file name (model.safetensors next to model.vae.pt) so a bad VAE never applies globally.
Never put --lowvram and --medvram together; the later one wins and the console warns.
The console line "Applying attention optimization: sdp" confirms the fast attention is active; if it says "Doggettx", add --opt-sdp-attention.
An SDXL checkpoint with a 1.5 LoRA produces a silent no-op or a size error; this site names the family on every model page, check it before you download.
Clear the browser cache (or open an incognito window) after an update; stale JavaScript produces strange UI errors that look like server faults.

Troubleshooting

Torch is not able to use GPU

Why it happens

CPU-only torch, old driver, or non-NVIDIA card.

How to fix it

Update the driver, delete venv and relaunch to reinstall torch with CUDA. AMD on Windows: use the DirectML fork or ZLUDA; on Linux use ROCm.

NansException with SDXL

Why it happens

fp16 VAE overflow.

How to fix it

Download sdxl-vae-fp16-fix, select it in SD VAE, or add --no-half-vae.

CUDA out of memory with hires fix

Why it happens

The upscaled pass needs several times the memory.

How to fix it

Lower the upscale factor, use a latent

upscaler

, switch Tiled VAE on, or run the hires pass as

img2img

afterwards.

Couldn't install torch

Why it happens

Python 3.11 or 3.12 instead of 3.10.

How to fix it

Install Python 3.10.6, point webui-user.bat at it with set PYTHON=path\python.exe, delete venv, relaunch.

Stable diffusion model failed to load

Why it happens

Half-downloaded or renamed file.

How to fix it

Re-download, check size against the model page, keep the original extension.

Extensions tab shows errors after git pull

Why it happens

Extensions written for an older WebUI.

How to fix it

Update each extension from its tab; remove the ones no longer maintained (the Manager-style list shows last commit dates).

Model loads but every image is noise

Why it happens

Wrong VAE applied globally, or a merged model needing a specific config.

How to fix it

Set SD VAE to Automatic, remove a global VAE, and put the checkpoint's .yaml next to it if the author provides one.

No install required - run AI models on the cloud

PirateDiffusion is Telegram only and built for pros: thousands of models, LoRAs and workflows driven by chat commands, with unlimited generation on a fixed price plan.

Try PirateDiffusion

Questions

Should I switch to Forge?

If you have 12 GB or less, yes; Forge runs SDXL and

Flux

where A1111 struggles, and your models and most extensions carry over. The Forge guide on this site lists the differences.

Can A1111 run Flux?

Not well. Flux support lives in Forge and

ComfyUI

. The Flux models on this site name the UIs they work with.

Where are my images?

In outputs/txt2img-images and outputs/img2img-images, dated folders. The Settings > Saving tab changes the pattern.

What do the model pages on this site give me for A1111?

The exact family (SD 1.5 or SDXL), trigger words, the file hash to check a download, and the Civitai source. LoRAs are called in the prompt as <lora:filename:0.8>.