Stable Diffusion WebUI Forge is the easiest way to get the classic Automatic1111 style interface with modern speed and memory handling. You type a prompt, pick a model from a dropdown and click Generate, with no node graphs to wire up.
There is one thing most older guides miss: the original Forge by lllyasviel has barely changed since late 2024, and the version people actually use for new models in 2026 is Forge Neo, a maintained fork by Haoming02.
Neo runs Flux, Qwen-Image, Z-Image, Krea 2, MOSS-Image, Anima, Chroma, Wan 2.2 video and more, on the latest PyTorch.
This guide shows you how to install Forge Neo on Windows, where every model file goes, which launch flags are worth using, how to share models with ComfyUI, and how to fix the most common errors. Facts come from the official Forge and Forge Neo READMEs and the Neo wiki, checked in November 2026.
- Which Forge: use Forge Neo (Haoming02/sd-webui-forge-classic, branch
neo). Original Forge is based on A1111 1.10.1 and its status table was last tested in 2024. - Install:
git clone https://github.com/Haoming02/sd-webui-forge-classic sd-webui-forge-neo --branch neo, thenuv venv venv --python 3.13 --seed, add--uv, runwebui-user.bat. - Opens at:
http://127.0.0.1:7860(change with--port). - Models: checkpoints in
models/Stable-diffusion, LoRAs inmodels/Lora, VAEs inmodels/VAE, text encoders inmodels/text_encoder, upscalers inmodels/ESRGAN. - Share ComfyUI models: add
--forge-ref-comfy-home "D:\ComfyUI". - RTX 50 series: do not use
--xformers(not supported); native PyTorch attention or--sageis fine.
What Is Forge WebUI?
Forge is a platform built on top of Automatic1111’s Stable Diffusion WebUI. Its author, lllyasviel (also the creator of ControlNet), made it to speed up inference, manage GPU memory better and test new features. The name comes from Minecraft Forge.
You get the same tabs A1111 users know: txt2img, img2img with inpainting, Extras for upscaling, PNG Info, a checkpoint merger and an Extensions tab. Under the hood, Forge swapped in a much better memory system, so models that crash A1111 often run fine in Forge on the same card.
ControlNet, IP-Adapter, FreeU, Hires. fix and X/Y/Z Plot are built in, so you need far fewer extensions than with plain A1111.
Original Forge vs Forge Neo
The original Forge README says it is based on SD-WebUI 1.10.1, and its own “Forge Status” table shows the last manual tests in August and September 2024. Its news file talks about a Gradio 5 upgrade planned for 2025. In practice, development slowed down a lot after that.
Forge Neo describes itself as a continuation of the “latest” Forge build. It focuses on optimisation, usability and running the newest popular models through the same easy GUI.
| Feature | Original Forge | Forge Neo |
|---|---|---|
| Repo | lllyasviel/stable-diffusion-webui-forge | Haoming02/sd-webui-forge-classic (branch neo) |
| Install | One-click 7z package (CUDA 12.1 + PyTorch 2.3.1) | git clone + uv or Python 3.13 |
| PyTorch | 2.1 to 2.4 | 2.14.0 + CUDA 13.0 |
| Flux | Yes (NF4, GGUF) | Yes (GGUF, fp8, nvfp4 and more; no bitsandbytes) |
| Qwen-Image, Z-Image, Krea 2, MOSS-Image, Anima | No | Yes |
| Wan 2.2 video | No | Yes (14B) |
| SD2 and SD3 | SD3 on a side branch | Removed |
| Best for | Old setups and SD1/SDXL only | Anyone starting fresh in 2026 |
If you only use SD1.5 or SDXL and already have the old one-click package working, there is no rush to move. For anything newer, install Neo.
Requirements
| Item | Minimum | Recommended |
|---|---|---|
| OS | Windows 10/11, Linux, macOS | Windows 11 or Linux |
| GPU | NVIDIA with 6 GB VRAM (SD1.5/SDXL) | NVIDIA with 12-16 GB for Flux, Qwen-Image, Wan |
| RAM | 16 GB | 32 GB or more (offloading uses system RAM) |
| Disk | About 10 GB plus models | SSD with 100 GB free |
| Software | git, Python 3.13 or uv | git + uv, FFmpeg for video export |
AMD users can run Neo through ROCm (the README links to CS1o’s guide), but the maintainer says AMD issues are not handled, so expect some trial and error.
Install Forge Neo on Windows
The recommended method uses uv, a very fast Python package manager. It makes the first install much quicker and can keep the folder small.
- Install git. Download it from git-scm.com and keep the default options.
- Install uv. Open PowerShell and run
powershell -ExecutionPolicy ByPass -c "irm https://astral.sh/uv/install.ps1 | iex", then close and reopen the terminal. - Clone Neo. Go to a drive with lots of space (for example
D:\) and run:git clone https://github.com/Haoming02/sd-webui-forge-classic sd-webui-forge-neo --branch neo - Create the Python environment.
cd sd-webui-forge-neouv venv venv --python 3.13 --seed - Add the uv flag. Open
webui-user.batin Notepad and change the args line toset COMMANDLINE_ARGS=--uv. Save the file. - Launch. Double-click
webui-user.bat. The first launch installs every requirement, which takes a few minutes. - Open the UI. When setup finishes, your browser opens
http://127.0.0.1:7860automatically.
Not keen on uv? The older method still works: install Python 3.13 with “Add Python to PATH” ticked, skip steps 2, 4 and 5, and run webui-user.bat. Linux and macOS users follow the Unix page on the Neo wiki, and there is a Docker folder for NVIDIA and AMD containers.
To update later, open a terminal in the folder and run git pull, then launch as usual.
Generate Your First Image
- Download an SDXL checkpoint (a
.safetensorsfile) and put it inmodels/Stable-diffusion. - In the UI, pick the xl preset at the top so the default sizes and settings match SDXL.
- Click the refresh icon next to the checkpoint dropdown and choose your model.
- Type a prompt such as a cosy reading nook with warm lamp light, rain on the window, detailed, soft focus.
- Set 1024 x 1024, 25-30 steps and CFG around 5-7, then click Generate.
Neo saves the checkpoint, text encoder and VAE choice plus parameters per preset, so switching between SDXL and Flux keeps the right settings for each. By default image sizes must be multiples of 64 (you can change this in Settings, System).
One UI tip from the original Forge team that still applies: in the Gradio 4 canvas, you move the canvas with the right mouse button.
Where Model Files Go
All folders live inside sd-webui-forge-neo\models\. Neo also moved the embeddings folder inside models.
| File type | Folder | Examples |
|---|---|---|
| Checkpoints and diffusion models | models/Stable-diffusion |
SDXL checkpoints, Flux, Qwen-Image, Z-Image (safetensors or GGUF) |
| Text encoders | models/text_encoder |
clip_l, t5xxl, gemma, Qwen encoders |
| VAE | models/VAE |
sdxl_vae, Flux ae.safetensors |
| LoRA | models/Lora |
Style and character LoRAs |
| Upscalers | models/ESRGAN |
All .pth and .safetensors upscalers (Neo puts every upscaler here) |
| ControlNet | models/ControlNet |
SDXL Union, LLLite |
| Embeddings | models/embeddings |
Textual inversion files |
For split models like Flux, you load the diffusion model as the checkpoint and pick the text encoders and VAE in the VAE / Text Encoder dropdown next to it. The Neo wiki’s Download Models page lists the exact encoder and VAE files for each architecture.
Some models are detected by name. A Qwen edit model needs both “qwen” and “edit” in its file or folder name, and a Flux Kontext model needs “kontext”. If an edit model behaves like a normal one, rename the file.
Supported Models in Neo
| Family | Notes |
|---|---|
| SD1.5 and SDXL | Including v-prediction, Zero Terminal SNR and rectified flow SDXL models |
| Flux.1 Dev, Krea, Kontext | bf16, fp8 or GGUF |
| Flux.2 Klein | 4B and 9B (not Flux.2 Dev) |
| Qwen-Image, Qwen-Image-Edit, Qwen-Image-2.1 | Edit models need “qwen” + “edit” in the name |
| Z-Image | Fast photoreal model |
| Krea 2 | Plus Identity Edit with a LoRA |
| MOSS-Image and MOSS-Image-Edit | Edit needs a specific LoRA |
| Anima 2B / 2.9B / 3.8B | Anime models, plus Anima Edit and RGBA |
| Chroma1-HD, Lumina-Image-2.0, Ernie-Image | Supported natively |
| Wan 2.2 14B | Text and image to video; FFmpeg needed to export |
Removed in Neo: SD2, SD3, hypernetworks, Forge Spaces, the CLIP and Deepbooru interrogators, textual inversion training and bitsandbytes (NF4) support. If you relied on NF4 Flux files, switch to GGUF or fp8 versions.
Not sure which model to start with? See our list of the best open source text-to-image models, and if you also use ComfyUI, the ComfyUI model folder cheat sheet shows where the same files go there.
Useful Launch Flags
Add flags to the set COMMANDLINE_ARGS= line in webui-user.bat, separated by spaces. Run python launch.py --help to see them all.
| Flag | What it does |
|---|---|
--uv |
Uses uv instead of pip for much faster installs |
--uv-symlink |
Same, but links packages from the cache, cutting install size from about 7 GB to about 100 MB |
--port 7861 |
Changes the port (default 7860) |
--api |
Turns on the txt2img/img2img API |
--sage |
Installs SageAttention (and tries Triton) for faster generation |
--cuda-malloc |
Better memory allocation |
--cuda-stream |
Async weight offloading |
--pin-shared-memory |
Better use of system RAM when offloading |
--fast-fp16 |
Enables fp16 accumulation for extra speed |
--pynvml |
Reads real free VRAM to avoid freezing other apps |
--model-ref "D:\models" |
Uses one central models folder instead of the built-in one |
A good starting line for an NVIDIA card is set COMMANDLINE_ARGS=--uv --cuda-malloc --pin-shared-memory.
Share Models With ComfyUI or A1111
Model files are huge, so do not copy them twice. Neo can read another app’s folders directly:
--forge-ref-comfy-home "D:\ComfyUI"loads ComfyUI’sdiffusion_models,clipand other folders.--forge-ref-comfy-yaml "D:\ComfyUI\extra_model_paths.yaml"reuses your existing ComfyUI path config.--forge-ref-a1111-home "D:\stable-diffusion-webui"loads an old A1111 install’s model folders.
Note that --model-ref replaces the models folder completely, while the forge-ref flags add the other app’s folders on top.
Speed and VRAM Tips
- Pick the right file size. On 8-12 GB cards use GGUF Q4 to Q6 or fp8 versions of big models. Our GGUF quantization guide explains which quant to pick.
- Do not stack attention packages. Neo picks SageAttention, then FlashAttention, then xformers, then PyTorch. The README warns that native PyTorch attention is usually as fast and more stable, so only add
--sageif you measure a gain. - Use fewer steps with turbo LoRAs. Many new models have distilled 4-8 step LoRAs.
- Enable live preview carefully. TAESD previews are cheap, but full previews slow things down.
- Close other GPU apps. Browsers with hardware acceleration and games eat VRAM. Neo also unloads all models after an out-of-memory error, so you can retry straight away.
Troubleshooting
| Problem | Likely cause | Fix |
|---|---|---|
| xformers errors on RTX 5070/5080/5090 | xformers does not support RTX 50 series | Remove --xformers; use PyTorch attention or --sage |
| “CUDA error: no kernel image” on an old GPU | Latest PyTorch dropped your card | Install an older PyTorch as shown on the Neo wiki’s Extra Installations page |
| Out of memory | Model too big for VRAM | Use a GGUF/fp8 file, lower resolution, add --cuda-malloc; try --expandable-segments |
| Black or grey images | Wrong or missing VAE / text encoder | Select the matching VAE and encoders from the wiki’s model table |
| Edit model ignores the input image | Name-based detection failed | Rename so the path contains “qwen” + “edit” or “kontext” |
| Video will not export | FFmpeg missing | Install FFmpeg and add it to PATH |
| Old A1111 extension crashes | Not built for Forge/Neo | Remove it; the maintainer does not support third-party extension issues |
| Install is slow or huge | pip downloads everything | Use --uv or --uv-symlink |
| Weird errors after a Stability Matrix install | Third-party launcher changes | Do a clean install with the official steps |
When something breaks after an update, the Neo maintainer only looks at issues you can reproduce on a clean install, so a fresh clone is often the fastest fix.
Forge vs ComfyUI
| Forge Neo | ComfyUI | |
|---|---|---|
| Interface | Simple tabs and sliders | Node graph |
| Learning curve | Low | Medium to high |
| New model support | Fast for popular models | Usually first, often on day one |
| Custom workflows | Limited to built-in features and extensions | Anything you can wire up |
| Best for | Quick prompting, inpainting, upscaling | Complex pipelines, video, automation |
Many people run both and share one models folder. Our ComfyUI beginner’s guide is the next step once you want more control.
My Take
If you liked A1111 but found it slow or crash-prone, Forge Neo is the version to install in 2026. It keeps the friendly interface, adds the modern models people actually want, and the uv install takes the pain out of Python setup.
Start with an SDXL or Z-Image model to confirm everything works, then add Flux or Qwen-Image once you know your VRAM limits. Want a click-to-install alternative? Pinokio can set up several image apps for you.
FAQ
Is Forge WebUI free?
Yes. Both original Forge and Forge Neo are free and open source. Model licences vary, so check each model card.
Is Forge still being updated?
The original lllyasviel Forge has had little development since late 2024. Forge Neo is the actively maintained fork with support for new models.
Does Forge support Flux?
Yes. Neo runs Flux.1 Dev, Krea and Kontext in bf16, fp8 and GGUF, plus Flux.2 Klein 4B and 9B. NF4 (bitsandbytes) files are no longer supported in Neo.
Can I use my A1111 extensions?
Many work, but not all. ControlNet and several common extensions are built in. Remove any extension that causes errors.
How much VRAM does Forge need?
About 6 GB for SD1.5 and SDXL. For Flux, Qwen-Image or Wan 2.2, 12-16 GB is comfortable, or use GGUF/fp8 files on smaller cards.
Can Forge use my ComfyUI models?
Yes. Add --forge-ref-comfy-home with your ComfyUI path, or point it at your extra_model_paths.yaml.
Does Forge make videos?
Forge Neo supports Wan 2.2 14B for video. You need FFmpeg installed to export the result.
Why is my RTX 50 card failing with xformers?
xformers does not support RTX 50 series cards. Remove the flag and let Neo use PyTorch attention.




