Fooocus in 2026: Is It Still Worth Installing, and What to Use Instead

For most people, no. Fooocus 2.5.5 can still turn a short prompt into a good SDXL image on an NVIDIA RTX 20, 30 or 40 card, but its last code change shipped on 12 August 2024. It cannot load FLUX, Qwen-Image or any other newer model family, it fails on RTX 50-series GPUs unless you swap PyTorch by hand, and its web server is pinned to a Gradio version with 42 published advisories. Install it only if you have a supported card, want SDXL and nothing else, and will keep it on your own machine. If you want the same one-box approach with current models, look at RuinedFooocus, FooocusPlus or SwarmUI first.

This guide covers what Fooocus still does well, the exact files it downloads, what fails today and the workarounds, the security position, which forks are alive, and how close each alternative gets to one simple screen. BFWR has not run Fooocus or any of these tools. Everything here comes from their READMEs, source code, issue trackers and published files, plus the GitHub, Hugging Face and advisory APIs, all read on 9 November 2026.

Quick answer

• Install it if: you have an RTX 20, 30 or 40 card, only want SDXL-family checkpoints, and value its inpainting and styles over new models.
• Skip it if: you have an RTX 50 card, want FLUX or anything newer, or would expose it on a network.
• Status: the README says “Limited Long-Term Support (LTS) with Bug Fixes Only”. The five commits since v2.5.5 changed only the readme and one CI file. The repository is not archived, and has 250 open issues and 66 open pull requests.
• Download: 2.0 GB for the Windows package, then 7.5 GB on first launch, so 9.5 GB before the first image. Inpainting and Image Prompt add more.
• Best replacements: RuinedFooocus if you want to stay in the Fooocus family, SwarmUI if you want a simple form on top of ComfyUI.

Maintenance status, from the API

The repository is lllyasviel/Fooocus, GPL-3.0, with about 53,800 stars. Its pushed_at date of 1 December 2025 is a branch push. On the main branch, the last commit that touched code is the v2.5.5 release of 12 August 2024. After that come two readme edits in August 2024, two more in January 2025, and a CI action bump on 2 September 2025. There have been no commits in the last 90 days.

The README states the policy plainly: “There are no current plans to migrate to or incorporate newer model architectures.” In September 2024, collaborator mashb1t described it as “in security maintenance only mode, so no new features.” The security section below shows that even that has not held.

What Fooocus still does well

The README promises that “the number of needed mouse clicks is strictly limited to less than 3” between download and first image, and the interface is a prompt box with Input Image and Advanced checkboxes. Beneath it:

• Prompt expansion and styles. A local GPT-2 model rewrites short prompts as the Fooocus V2 style. The six files in sdxl_styles/ define 277 more styles.
• Performance modes. Quality runs 60 steps, Speed 30, Extreme Speed 8, and Lightning and Hyper-SD 4 each, with the faster modes loading a LoRA automatically.
• Inpaint and outpaint. Fooocus uses its own patch, inpaint_v26.fooocus.patch, rather than a dedicated inpainting checkpoint. It applies to ordinary SDXL checkpoints.
• Image Prompt. Four modes: ImagePrompt, PyraCanny, CPDS and FaceSwap.
• Input Image tools. Vary (Subtle or Strong), Upscale (1.5x, 2x, Fast 2x), Describe, and an Enhance tab with automatic masking.

The low-memory claim is the other draw: “Minimal GPU memory requirement is 4GB (Nvidia)”, with 8GB of system RAM. That depends on swap. The troubleshooting page says “Please make sure that you have at least 40GB System Swap.” The README’s own minimums:

What it cannot do

This is in the code, not only the README. The model list in ldm_patched/modules/supported_models.py is Stable_Zero123, SD15, SD20, SD21UnclipL, SD21UnclipH, SDXLRefiner, SDXL, SSD1B, Segmind_Vega, SD_X4Upscaler and SVD_img2vid. There is no class for FLUX, FLUX.2, FLUX.3, Stable Diffusion 3, Qwen-Image, Z-Image or Krea, and a search of every Python file for flux or sd3 returns nothing. For what the newer models need, see our local image model comparison.

Files, folders and download size

Each preset names one checkpoint, saved to Fooocus\models\checkpoints. Sizes are from the Hugging Face file listings, in decimal GB.

On Windows the real total is the 2.00 GB Fooocus_win64_2-5-0.7z, then 7.11 GB for Juggernaut, 0.05 GB for sd_xl_offset_example-lora_1.0.safetensors in models\loras, 0.35 GB for the prompt expansion model in models\prompt_expansion\fooocus_expansion, and 0.006 GB of preview files in models\vae_approx: 9.51 GB before the first image.

The smallest workable setup reuses an SDXL checkpoint you already own. Set path_checkpoints and default_model in config.txt, then launch with --disable-preset-download and no --preset, because a preset overrides config.txt. The expansion and preview files still download, so the total is about 2.36 GB on Windows.

Other features download on first use:

How to install it

Windows: download Fooocus_win64_2-5-0.7z from the v2.5.0 release, extract it, and run run.bat. That is still the current Windows package; the launcher, entry_with_update.py, then fast-forwards the code to main.

Only use GitHub. The README says “Do NOT get Fooocus from those websites”, naming fooocus.com, .net, .co, .ai, .org, .pro and .one as fakes.

Linux and Apple Silicon use the same steps, and need Python 3.10:

git clone https://github.com/lllyasviel/Fooocus.git
cd Fooocus
conda env create -f environment.yaml
conda activate fooocus
pip install -r requirements_versions.txt
python entry_with_update.py

On Linux with AMD, the README replaces torch with the rocm5.6 build. On Mac, it calls the steps an unofficial guideline and says “Mac is not intensively tested.”

The failures people actually hit

RTX 50-series: no kernel image

This is the big one, and it is spread across at least seven issues, including #3862 (81 comments), #4123 and #4222 from August 2026. The app starts, then fails at generation:

NVIDIA GeForce RTX 5050 Laptop GPU with CUDA capability sm_120 is not compatible with the current PyTorch installation.
RuntimeError: CUDA error: no kernel image is available for execution on the device

The cause is launch.py, which installs torch==2.1.0 from the CUDA 12.1 index. As mashb1t wrote in #3862, “turns out pytorch 2.1.0 isn’t compatible with 5000 series GPUs and has to be updated.” A fix, pull request #4155, was closed unmerged in July 2026.

The workaround users report is replacing PyTorch inside the bundled Python. This one, posted in #4123 and confirmed there on an RTX 5070 and 5070 Ti, runs from the extracted folder on Windows:

python_embeded\python.exe -m pip install --upgrade --force-reinstall --pre torch torchvision torchaudio --index-url https://download.pytorch.org/whl/nightly/cu128

Others in #3862 reported success with the stable whl/cu128 index. It is unsupported: mashb1t warned “it’s going to be a bumpy ride”. An easier route is Stability Matrix, whose Fooocus.cs installer detects these cards and selects the CUDA 12.8 build.

Other common errors

Security: keep it on your own machine

Defaults. Fooocus binds to 127.0.0.1 on port 7865. --listen with no address binds 0.0.0.0, and --share creates a public .gradio.live link. The README is direct: “In both ways the access is unauthenticated by default.” Basic login needs an auth.json file. If you use the Docker command in docker.md, change -p 7865:7865 to -p 127.0.0.1:7865:7865. Add --disable-analytics if you want Gradio’s analytics off.

Old dependencies. requirements_versions.txt pins gradio==3.41.2. GitHub’s advisory database lists 42 advisories affecting that version: 3 critical, 18 high, 18 medium and 3 low. That count is an upper bound: two are marked as duplicates and not all will be reachable in Fooocus. But one, CVE-2024-1727, “covers the ability of 3rd party websites to access routes and upload files to users running Gradio applications locally”, so a localhost bind does not cover everything. The database also returns 13 entries for torch 2.1.0 and 21 for pillow 10.4.0. Python 3.10 itself is now out of support: PEP 619 says “security updates (source only) were released until 1st October 2026”.

Unresolved reports. Issue #4115 reports code execution through image metadata. The code matches: modules/meta_parser.py calls eval() on values read from an image’s metadata when you use Input Image, Metadata, Apply Metadata. #4185 makes the same claim in more detail, and #4198 covers Gradio’s /file= route. These are third-party reports with no maintainer reply. GitHub Security Lab asked for a contact in August 2024 and wrote that October, “We have tried to reach out to your email but have not received a response.” The practical rule: do not apply metadata from images you did not make.

One more: launch.py sets ssl._create_default_https_context = ssl._create_unverified_context, which turns off certificate checks for standard-library HTTPS, and model downloads go through torch.hub, which uses it.

Community forks, by commits rather than stars

The 90-day window starts on 11 August 2026.

RuinedFooocus is the clear lead: of these, only it has moved to a current Gradio. It installs from rf-setup.ps1 on Windows or rf-setup.sh on Linux. FooocusPlus runs non-SDXL models through an integrated Comfy backend and says it “does support NVIDIA Blackwell 50xx graphics cards in full native mode”, but keeps a Gradio 3 pin. Avoid SimpleSDXL: Stability Matrix blocks it with a critical advisory, “Undisclosed Data Collection and Remote Access in simpleai_base Dependency”, and Fooocus removed its link in January 2025.

What to use instead

None of these is quite as bare as Fooocus. Our image editor roundup compares them as editors; this is how each serves someone leaving Fooocus.

If you want the simplest swap, SwarmUI’s README says “Beginner users will love Swarm’s primary Generate tab interface”, and its model table runs from SDXL to Flux 2, Flux 3 Image, Qwen Image 2.1, Z-Image and MOSS Image.

If you came for inpainting, Fooocus’s patch outlives it. ComfyUI Inpaint Nodes loads inpaint_v26.fooocus.patch from ComfyUI/models/inpaint and applies it to SDXL checkpoints, and Krita AI Diffusion lists the same two files as required resources for SDXL.

If you want to keep Fooocus available, Stability Matrix still installs it, along with four variants, though its docs file them under packages “no longer actively maintained or have been superseded by newer alternatives.”

Your checkpoints come with you: they are ordinary SDXL .safetensors files (see our checkpoint guide). To see what your card can hold, try the image generation VRAM calculator.

What about Forge?

Fooocus’s README recommends WebUI Forge for Flux, but the original Forge has stalled too: last commit on main 26 June 2025, none in 90 days. Its maintained continuation is Forge Neo, on the neo branch of Haoming02/sd-webui-forge-classic, with 148 commits in 90 days and release 2.31 on 21 October 2026. It is a many-control AUTOMATIC1111-style interface, not a Fooocus replacement.

The licences

Fooocus itself is GPL-3.0. Running it privately carries no obligations, and it says “The output from running a covered work is covered by this License only if the output, given its content, constitutes a covered work.” If you distribute a modified copy, you must offer the source. RuinedFooocus and FooocusPlus are GPL-3.0 too, and FooocusPlus adds its own usage conditions.

The model is a separate licence. Juggernaut XL v8 is tagged CreativeML OpenRAIL-M, and RunDiffusion’s card adds that it “may not be deployed behind paid API services” without a licence from them. Fooocus downloads it from a Hugging Face mirror whose README reads “This space is for my personal use only, not a distributing page.” The Fooocus inpaint repository is tagged OpenRAIL. Check our model licence checker before commercial use. This is our reading of the published texts, not legal advice.

FAQ

Is Fooocus dead?

It is not archived and it still works on supported hardware, but no code has changed since August 2024 and security reports go unanswered. Treat it as frozen.

Can Fooocus run Flux?

Not the official version; its loader has no Flux support. RuinedFooocus and FooocusPlus do list Flux.

Does Fooocus work on an RTX 5060, 5070, 5080 or 5090?

Not as shipped. It needs a newer PyTorch, installed by hand or by Stability Matrix’s Fooocus package. Users report it working afterwards, but nothing upstream supports it.

How much VRAM does Fooocus need?

The README says 4GB on RTX 2000 cards or newer, with 8GB of RAM and generous swap, and 8GB on GTX 1000 and AMD cards.

Related guides and tools

• SwarmUI: setup and models
• Stability Matrix: one installer for local image tools
• ComfyUI install: Desktop, Portable or Manual
• ComfyUI templates
• Stable Diffusion image editors
• Best local image generation models
• Image generation VRAM calculator
• Image model directory

Sources

• lllyasviel/Fooocus: README, troubleshoot.md, launch.py, requirements_versions.txt, presets, modules/config.py, modules/meta_parser.py, ldm_patched/modules/supported_models.py, and the issues linked above
• Hugging Face file listings: lllyasviel/fav_models, mashb1t/fav_models, lllyasviel/misc, mashb1t/misc, lllyasviel/fooocus_inpaint and the Juggernaut XL v8 card
• GitHub Advisory Database: CVE-2024-1727 and CVE-2023-51449, plus API queries for gradio 3.41.2, torch 2.1.0 and pillow 10.4.0
• PEP 619, Python 3.10 release schedule
• Stability Matrix Fooocus.cs and supported packages
• Repositories and READMEs read through the GitHub API: the forks and tools in the tables above, SwarmUI, ComfyUI and Forge

Updated 9 November 2026. The Fooocus source (main, commit ae05379 of 2 September 2025), its issues, the file sizes, the advisory counts and the status of every fork and alternative were read on that date.