.ai

de bästa AI-modellerna med öppen källkod

Sök

Stable Diffusion WebUI Forge errors and fixes (Flux, SDXL, low VRAM)

Q: AttributeError: NoneType ... when switching models

UI selector in the wrong mode. Set sd, xl or flux to match the checkpoint, then reselect the model.

Q: ControlNet from A1111 crashes on start

The A1111 extension hooks the old backend. Delete it from extensions; use the built-in ControlNet tab and its models in models/ControlNet.

Ämne: Fel och lösningar

Av Captain

Publicerad 2026-05-16

Forge-specific problems and their fixes: Flux not loading or producing noise, the UI type switch (sd / xl / flux), "GPU Weights" slider and out-of-memory errors, missing text encoders and VAE for Flux, ControlNet and extension breakage, and the differences from Automatic1111 that trip people up.

Den här sidan har inte översatts ännu, så den visas på engelska.

Översikt

Forge is lllyasviel's fork of the

Automatic1111

WebUI built on a new backend with aggressive memory management. It runs

Flux dev

on 6 to 8 GB cards and

SDXL

on 4 GB, which is why so many people migrate to it. The price is a set of errors that A1111 users have not seen: a UI mode switch that must match the model, a GPU Weights slider that decides whether a model fits, and separate files for

Flux

(diffusion model, two text encoders, VAE) that must all be present.

The rule that fixes half of all Forge problems: the UI selector at the top left (sd, xl, flux, all) must match the

checkpoint

, and for Flux the VAE / Text Encoder box must list ae.safetensors, clip_l.safetensors and a t5xxl file. The other half is the GPU Weights value: too high and you get

out of memory

, too low and generation is slow but works.

Hosted alternatives for what still does not fit:

PirateDiffusion

serves Flux, SDXL and the newer families through hosted workflows from Telegram, and

BitVector

runs the same models in a web app and on Discord. The model pages on this site carry a Run box for both.

Referens

Namn	Typ	Vad det är
UI selector (sd / xl / flux / all)	setting	Chooses the backend mode and which pulldowns show. A Flux checkpoint under "xl" fails with a shape or NoneType error.
GPU Weights (MB)	setting	How much of the model stays on the GPU. Leave about 1.5 to 2 GB free for activations: on a 12 GB card set 9000 to 10000 for Flux.
Swap Method / Swap Location	setting	Queue vs Async, CPU vs Shared. Async plus Shared is fastest on Windows with a modern driver; Queue plus CPU is the safe fallback.
Diffusion in Low Bits	setting	Automatic picks fp8 for Flux. Set it to bnb-nf4 for the NF4 checkpoints or fp8_e4m3fn for the fp8 ones; a mismatch produces noise.
VAE / Text Encoder	setting	For Flux select ae.safetensors, clip_l.safetensors and t5xxl_fp16 or t5xxl_fp8_e4m3fn. Missing t5 gives blank or garbage images.
models/text_encoder, models/VAE	folders	Where Forge looks for Flux encoders and VAE. A1111 users often put them in the checkpoint folder by mistake.
CUDA out of memory / Allocation on device	error	GPU Weights too high for the resolution and batch. Lower it by 1000 MB steps.
NoneType has no attribute ... in forge_loader	error	Wrong UI mode for the model, or a LoRA from another family.
Extension errors on start	error	A1111 extensions that hook the old backend (some ControlNet forks, Adetailer versions, Dynamic Prompts) break on Forge. Use the Forge-compatible versions.

Inget att installera - kör AI-modeller i molnet

BitVector Prism är det enklaste sättet att komma igång: välj en modell, skriv en prompt och generera i en ren webbapp, utan någon installation alls. BitVector finns också på Discord och på webben (SpyGlass).

Prova BitVector

Steg för steg

Install from the official one-click package or git clone; run update.bat once, then run.bat. Forge bundles its own Python and torch.
Put files in the Forge folders: checkpoints and Flux diffusion models in models/Stable-diffusion, text encoders in models/text_encoder, VAEs in models/VAE, LoRAs in models/Lora. Forge also reads an A1111 folder tree through webui-user.bat --ckpt-dir and --lora-dir.
Switch the UI selector to the family of the model you load. For Flux dev fp8 select flux, then in VAE / Text Encoder tick ae.safetensors, clip_l.safetensors and t5xxl_fp8_e4m3fn.safetensors.
Set GPU Weights: total
VRAM
in MB minus about 2000 for 1024 pixel images, minus 3000 if you use hires fix or ControlNet. Generate once; if it fails with out of memory, lower by 1000 and repeat.
Leave Diffusion in Low Bits on Automatic unless you downloaded an NF4 checkpoint (then bnb-nf4). The Flux model pages on this site name the quantisation in the file list.
Flux settings that differ from SDXL: Distilled
CFG
Scale 3.5 and CFG 1, 20 to 28 steps, Euler with Simple schedule.
Negative prompts
are ignored at CFG 1.
If extensions fail, start with --disable-all-extensions, then enable the Forge-maintained versions (built-in ControlNet, built-in FreeU and Kohya HRFix are already there).
For video or models beyond Forge (
Wan
,
Qwen-Image
,
HunyuanVideo
), move to
ComfyUI
or run them hosted on BitVector or PirateDiffusion; Forge is an image UI.

Exempel

webui-user.bat sharing an A1111 model folder

set COMMANDLINE_ARGS=--ckpt-dir D:\sd\models\Stable-diffusion --lora-dir D:\sd\models\Lora --vae-dir D:\sd\models\VAE

Flux dev fp8 settings

UI: flux | Checkpoint: flux1-dev-fp8.safetensors | VAE/Text Encoder: ae.safetensors, clip_l.safetensors, t5xxl_fp8_e4m3fn.safetensors

Tips

Forge ships ControlNet, FreeU, Kohya HRFix, SAG and more as built-ins; remove the A1111 versions from the extensions folder before migrating.
The NF4 Flux checkpoint (about 6.8 GB with encoders baked in) is the simplest Flux on 8 GB cards; the fp8 version with separate encoders gives slightly better detail on 12 GB.
Set Swap Method to Async and Swap Location to Shared, generate, and switch back to Queue/CPU if the driver stutters. Async can double speed on 8 GB cards.
The seed, sampler and CFG from an A1111 PNG reproduce on Forge for
SD 1.5
and SDXL; Flux images need the Distilled CFG value too.
Keep LoRA weights for Flux at 0.6 to 0.9; Flux LoRAs saturate faster than SDXL ones. Trigger words are on each model page here.
Forge logs the memory plan in the console ("Moving model(s) has taken...") - watch it once to understand what the GPU Weights slider does.

Raise CFG to 2 to 3 (slower, and colours shift) or describe what you want positively.

Inget att installera - kör AI-modeller i molnet

PirateDiffusion finns bara på Telegram och är byggt för proffs: tusentals modeller, LoRA och arbetsflöden som styrs med chattkommandon, med obegränsad generering till ett fast pris.

Prova PirateDiffusion

Frågor

Does Forge replace A1111 completely?

For SD 1.5, SDXL and Flux images, yes, with less memory and a faster backend. A few niche extensions never moved over. Video and audio models need ComfyUI.

Which Flux file should I download?

On 6 to 8 GB: the NF4 checkpoint. On 12 GB: fp8 with separate encoders. On 24 GB: the full fp16 dev. The Flux family page on this site lists the base files and the LoRAs.