RuinedFooocus in 2026: Models, Features, and Setup

Current Status — June 2026

RuinedFooocus is active. Its current source version is 26.06.27, with June updates for Krea 2, Boogu Image, Gradio 6, and runtime changes.

The confusing part is the download page. GitHub's latest tagged Windows archive is still Release 2.0.0 from January 2025, while the main branch has moved far beyond it.

Use the tagged archive for the simpler Windows install. Use the current source if you need the newest model support. Those aren't the same build, so check the update log before troubleshooting a missing feature.

What RuinedFooocus Actually Is Now

Current RuinedFooocus is a prompt-first local generator built from Fooocus roots, with pinned ComfyUI and ComfyUI-GGUF components underneath. It keeps the familiar browser interface but supports far more than official Fooocus.

That distinction matters. Official Fooocus stays focused on SDXL and receives bug fixes only. RuinedFooocus has continued adding model loaders, image tools, and video pipelines.

If you want the short version:

• Official Fooocus: Stable SDXL tool with a narrow, mature feature set.
• RuinedFooocus: Broader community project for newer image and video architectures.
• ComfyUI: Better when you need reusable graphs and complete pipeline control.
• Forge: Better when you want a conventional WebUI with a larger extension ecosystem.

Read the Fooocus vs RuinedFooocus comparison if you're choosing between the two interfaces.

Model Support

RuinedFooocus's model list has grown well beyond SDXL.

• Stable Diffusion: SDXL, SD 1.5, SD3, and SD3 Large support appear in the official update history.
• Flux: Flux.1, Flux.1 Kontext, Flux.2 Dev, GGUF files, and UNet-only safetensors are documented.
• Newer image models: AuraFlow, Qwen Image, HiDream, Z-Image Turbo, Anima, PixelDiT, Ernie Image, Ideogram 4, Boogu Image, and Krea 2 appear in recent updates.
• Video: The current repository contains WAN, Hunyuan Video, LTX Video, and LTX2 pipeline modules.
• LoRAs and local files: The UI supports LoRA selection, trigger-word lookup, thumbnails, subdirectories, and custom model paths.

This doesn't mean every random checkpoint will load. Match the model architecture to a supported loader, then check the update log for model-specific limitations.

The Features That Still Matter

The broad model list gets attention, but the workflow tools are the reason to use RuinedFooocus instead of a bare loader.

• Prompt batching: Separate prompts with --- and process a list without submitting each job manually.
• Image controls: ControlNet, image-to-image, inpainting, upscaling, background removal, and interrogation are documented.
• Prompt helpers: Wildcards, styles, random prompts, automatic negative prompts, and inline LoRA syntax reduce repetitive setup.
• Model organization: Subfolders, previews, metadata viewing, and trigger-word files help when your model folder gets crowded.
• Video experiments: WAN, Hunyuan, and LTX pipelines put image and video tests in one interface.
• Offline launch mode: Once dependencies and models are installed, the launcher can skip update and package checks with offline mode.

One old feature needs a correction: MergeMaker isn't a safe current selling point. The README still lists it, but the update log says merge management was removed in June 2025. Current code search doesn't show the old MergeMaker implementation.

Installation: Pick the Right Path

Windows tagged release

1. Download the official .7z archive from the project's GitHub releases.
2. Extract it with a current version of 7-Zip.
3. Run run.bat.
4. Let the first launch download its default SDXL model.

This is the shortest path, but the tagged archive is older than the current source.

Current source on Linux or Windows

1. Clone runew0lf/RuinedFooocus from GitHub.
2. Create and activate a virtual environment.
3. Install requirements_versions.txt.
4. Run python launch.py.
5. Add the models needed for your selected architecture.

The current launcher detects the hardware platform and installs matching Torch packages through torchruntime. It also pulls pinned runtime components unless you start in offline mode.

Don't use random download mirrors. The official repository and release page are the only sources covered by the project documentation.

Hardware Depends on the Model

There is no honest universal “RuinedFooocus needs X GB VRAM” answer. SDXL, Flux, SD3, Krea 2, and video models don't have the same memory requirements.

The official project doesn't publish one current minimum that covers every loader. Treat any page promising that all features run on 4 GB or 6 GB as outdated.

Choose your hardware target by architecture:

• SDXL workflows: Start with the requirements for the specific SDXL checkpoint.
• Flux and GGUF: Quantization changes memory use, quality, and speed.
• Video models: Expect much higher memory and storage demands than basic image generation.
• CPU mode: Supported in the update history, but that doesn't make heavy models practical on every CPU.

Unrestricted Workflows

RuinedFooocus is commonly used for unrestricted local generation. Current code search doesn't show the optional Fooocus NSFW blackout implementation in this repository.

That isn't a promise that every model follows every prompt. Check the model card, license, and training limits for each checkpoint you download.

If you only need uncensored SDXL and don't need newer architectures, regular Fooocus may already cover the job. Its blackout option is disabled by default in current official source. See the Fooocus uncensored setup guide before switching tools.

What It Gets Wrong

• Documentation drifts: The README and current code don't always agree. MergeMaker is the clearest example.
• Release labeling is confusing: The latest tagged archive is much older than the current branch.
• Setup grows with model support: More loaders mean more dependencies, model files, and architecture-specific errors.
• Community support is smaller: RuinedFooocus has fewer users and fewer tutorials than Fooocus, Forge, or ComfyUI.
• No universal benchmark: Exact generation times depend on the model, quantization, GPU, resolution, and settings.

Who Should Use RuinedFooocus?

Pick RuinedFooocus if you want newer models without building ComfyUI graphs. It makes the most sense for curious users who accept some troubleshooting in exchange for broader support.

Skip it if you only need stable SDXL generation. Official Fooocus is simpler and better documented for that narrow job.

Use ComfyUI if repeatable graphs matter more than interface simplicity. Use Forge if you want a conventional Stable Diffusion WebUI. LocalForge AI is another paid option if you want a preconfigured local setup instead of maintaining Python packages yourself.