Fooocus runs NSFW Pony, SDXL and Illustrious checkpoints locally with no content filter, because all generation happens on your own GPU. Install it from the official GitHub release, drop your checkpoints into the models/checkpoints folder, pick the model in the Advanced panel, disable prompt expansion for anime models, and generate. You need about 4GB VRAM minimum, 6GB or more recommended.
Why Fooocus for local NSFW
Fooocus by lllyasviel is the easiest way to run Stable Diffusion locally. It strips away the intimidating settings wall of other UIs and gives you a near one-click experience that still produces high-quality SDXL output. For NSFW creators, the headline benefit is simple: local generation has no content filter at all. There is no platform policy, no Stamina meter, no Buzz cost, and nothing phoning home to moderate your prompts. You own the GPU, you own the output.
Unlike cloud tools such as SeaArt or Holara that gate adult content, a local Fooocus install with a Pony or Illustrious checkpoint generates exactly what you prompt. This guide walks the full setup: install, where checkpoints go, presets, removing the few default behaviors that get in your way, common errors, VRAM needs, and how Fooocus stacks up against Forge and ComfyUI.
Step 1: Install Fooocus
Fooocus is Windows-friendly and needs no Python or Git knowledge for the standard route.
- Go to the official Fooocus GitHub and find the latest release in the readme.
- Download the one-click installer package (a 7z archive for Windows).
- Extract it with 7-Zip to a folder with plenty of free disk space. Checkpoints are large, so budget 30GB or more.
- Run run.bat. On first launch Fooocus downloads its default model and dependencies automatically, which takes a while.
- When it finishes, your browser opens to the Fooocus interface at a local address. That is it; you are running.
Nvidia GPUs get the smoothest experience. AMD and Mac (Apple Silicon) are supported with specific launch flags noted in the GitHub readme, though performance is slower. If you have an Nvidia card with at least 4GB VRAM, the default install just works.
Step 2: Where to put Pony, SDXL and Illustrious checkpoints
This is the core of NSFW setup. The default model Fooocus downloads is general-purpose. For adult work you want specialized checkpoints, and they go in one specific folder.
Place checkpoint .safetensors files here:
Fooocus\models\checkpoints\Download your checkpoints from Civitai, enabling its mature-content toggles first as covered in our Civitai NSFW generator guide. Good starting picks:
- Pony Diffusion V6 XL and its fine-tunes for stylized and anime NSFW. Learn the score-tag syntax in our Pony Diffusion guide.
- Illustrious XL fine-tunes for clean, high-detail anime. See our Illustrious XL guide.
- A realistic SDXL fine-tune for photoreal adult work. Browse our best NSFW checkpoints list.
LoRAs go in a sibling folder:
Fooocus\models\loras\Grab reliable ones from our best NSFW LoRAs roundup. After dropping files in, restart Fooocus (or refresh), and your new checkpoints appear in the Base Model dropdown inside the Advanced panel. Select your Pony or Illustrious checkpoint there before generating.
Step 3: Presets and the Advanced panel
Fooocus hides complexity behind presets. Tick the Advanced checkbox to reveal the controls that matter for NSFW work.
- Base Model – choose your downloaded Pony, Illustrious or SDXL checkpoint here. This overrides the default.
- LoRA slots – add LoRAs and set their weights.
- Performance – Speed, Quality, and the faster Extreme Speed and Lightning modes. Quality gives the best detail; the speed modes trade fidelity for fast iteration.
- Aspect ratio and image number – set resolution and batch size.
- Styles – Fooocus style presets. For Pony and Illustrious, turn most of these OFF, because they inject SDXL-style language the anime models do not respond to well.
Fooocus 2.5.0 and later added a dedicated Pony preset and Pony-friendly handling, so if your version offers a Pony preset, selecting it sets sensible defaults. You can switch presets in the browser without restarting.
Ready to test prompts before you commit to a full local install? Our free browser generator runs uncensored with no setup, so you can validate your prompt ideas first.
Step 4: Avoid the built-in behaviors that hurt NSFW output
Fooocus has no hard NSFW safety filter blocking local generation, which is the whole point of going local. But it has two default behaviors that interfere with adult anime models, and you should turn them down.
Prompt expansion (GPT-style). By default Fooocus rewrites and expands your prompt with extra descriptive tokens via its built-in expansion. This helps casual SDXL prompts but actively hurts Pony and Illustrious, which expect precise booru and score tags, not flowery prose. In the Advanced panel under Performance and styles, disable or minimize prompt expansion (turn off the Fooocus V2 expansion style) so your tags reach the model intact.
Default style injection. As noted, the default Fooocus styles add language tuned for the base SDXL model. For anime checkpoints, clear them so your prompt is exactly what you typed.
With those two off, your Pony or Illustrious checkpoint behaves the way it does on Civitai or a local Forge install, responding cleanly to your tags. There is no separate censorship toggle to flip; the model and your prompt do the work.
Step 5: VRAM needs and performance
Fooocus is notably efficient, which is why it runs on modest hardware.
| VRAM | Experience | Notes |
|---|---|---|
| 4GB | Minimum, works | Slow, use Extreme Speed mode |
| 6GB | Usable | Fooocus optimizes well here |
| 8GB | Comfortable | Good SDXL and Pony speed |
| 12GB (RTX 3060) | The sweet spot | Fast, headroom for LoRAs |
| 16GB+ | Excellent | Big batches, high res, fast |
The official minimum is 4GB Nvidia VRAM, with 6GB or more recommended for reasonable speed. The RTX 3060 12GB remains the value floor for serious local generation: cheap, 12GB of VRAM, and fast enough for daily Pony and Illustrious work. On 4GB to 6GB cards, lean on the Extreme Speed performance mode and smaller resolutions, then upscale keepers. Our best local NSFW generator guide covers hardware choices in depth.
Common errors and fixes
“CUDA out of memory.” Your VRAM is exhausted. Lower the resolution, reduce batch size to 1, switch to a faster performance mode, and close other GPU apps. Fooocus already offloads aggressively, so dropping resolution is the fastest fix.
New checkpoint does not appear in the dropdown. It is in the wrong folder or Fooocus did not rescan. Confirm the file is in models\checkpoints, then click the refresh icon next to the model dropdown or restart run.bat.
Anime prompts produce mushy or generic results. Prompt expansion is still on, or you are using SDXL-style sentences. Disable expansion, use proper booru and score tags for Pony and Illustrious, and select the correct base model.
Install fails or run.bat closes instantly. Usually a path issue or antivirus quarantine. Extract to a short path with no special characters, whitelist the folder in your antivirus, and ensure you fully extracted the archive rather than running it from inside the zip.
Slow first generation, then normal. Normal behavior. The first run loads the model into VRAM; subsequent generations are much faster.
Using LoRAs, inpainting and upscaling in Fooocus
Fooocus does more than basic text-to-image, and the extra features matter for polished NSFW work. LoRA support lets you stack character, style and concept LoRAs on top of your base checkpoint, controlling each one’s weight from the Advanced panel. Keep total LoRA weight reasonable, since stacking too many at full strength muddies the output and fights the base model. A character LoRA at moderate weight plus one style LoRA is usually plenty.
Fooocus also includes an Input Image section that unlocks inpainting, outpainting and image prompting. Inpainting lets you mask a region and regenerate just that part, which is the cleanest way to fix a hand, adjust an outfit, or correct anatomy without rerolling the whole image. Outpainting extends the canvas beyond its original borders, useful for reframing a composition. Image prompt and the Enhance feature let you guide generation from a reference image, which helps with character consistency across a set. None of these features carry a content filter; they operate on whatever your checkpoint produces.
For upscaling, Fooocus offers an Upscale option in the Input Image panel that takes a finished generation to higher resolution with added detail. The smart workflow is to generate and refine at native resolution, pick your keeper, then upscale only that one, rather than generating everything at high resolution from the start. This keeps generation fast and your VRAM happy, especially on 6GB to 8GB cards where high native resolutions risk out-of-memory errors.
Fooocus vs Forge vs ComfyUI
All three run the same checkpoints locally with no filter. They differ in how much control and complexity they expose.
Fooocus is the easiest. Minimal settings, smart defaults, great for getting clean SDXL, Pony and Illustrious output fast without learning a node graph. The trade-off is less granular control. It is the best starting point for most NSFW creators.
Forge sits in the middle. It is an optimized Automatic1111 fork with the full WebUI feature set: every sampler, ControlNet, extensions, and fine control, while running faster and lighter than vanilla A1111. Step up to it when you outgrow Fooocus. See our Forge NSFW setup guide.
ComfyUI is the most powerful and the steepest learning curve. Its node-based workflow gives total control over every step and is the choice for complex pipelines, custom workflows and advanced ControlNet stacks. Our ComfyUI NSFW guide walks through it.
The practical path: start with Fooocus, move to Forge when you want full control, graduate to ComfyUI when you need custom pipelines. For browser convenience without any install, compare cloud options in our best NSFW AI generators ranking, or just use our free generator.
Prompting Pony and Illustrious in Fooocus
The biggest mistake new Fooocus users make with anime checkpoints is prompting them like the default SDXL model. Pony and Illustrious are tag-trained, so they want structured tags, not descriptive sentences. For Pony Diffusion, open your prompt with the score tags it expects, then add Danbooru-style tags for the character, pose, outfit, expression and setting, ordered from most to least important. For Illustrious, drop the score prefix and use clean booru tags plus a short quality block, since it reacts better to concise tagging. In both cases, make sure prompt expansion is off so Fooocus does not bury your tags under generated prose.
Negative prompts matter just as much locally as they do in the cloud. A baseline negative that suppresses extra limbs, fused or extra fingers, low quality, blurriness, watermarks and text artifacts cleans up most output before you reroll. Many checkpoint pages on Civitai publish a recommended negative prompt, and pasting it into Fooocus is a fast way to match the look the model author intended. Combine that with a sensible step count, generally 25 to 35 steps in Quality mode, and you will get clean results without wasting generation time on diminishing returns.
Workflow tips for efficient local generation
Local generation is free per image, but your time and VRAM are not, so a few habits keep you productive. Use a faster performance mode like Speed or Extreme Speed for the exploration phase, generating small batches to find a composition and prompt you like, then switch to Quality mode and a higher resolution only for the final render. This is the local equivalent of not burning cloud credits on high-res rerolls, except here the cost is your time and patience rather than money.
Keep your checkpoints and LoRAs organized. As your collection grows, the Base Model and LoRA dropdowns get crowded, so name files clearly and prune ones you never use. When you find a prompt, checkpoint and settings combination that works, save the prompt text somewhere outside Fooocus, because reproducing a great result months later is painful if you only remember it vaguely. Fooocus also writes generation metadata into the output images, so you can drag an old image back in to recover its settings if you kept the file.
Finally, keep Fooocus updated. The project ships improvements and new model support regularly, and the Pony preset and anime handling that arrived in version 2.5.0 are a good example of why staying current pays off. Updating is usually as simple as running the update script included in the install package, and it rarely breaks an existing setup.
The bottom line on Fooocus NSFW setup
Fooocus is the friendliest on-ramp to uncensored local generation in 2026. Install the one-click package, drop Pony, SDXL and Illustrious checkpoints into models\checkpoints, select your model in the Advanced panel, disable prompt expansion and default styles for anime models, and you have a filter-free studio on your own GPU. With 6GB or more VRAM (a 12GB RTX 3060 is the sweet spot) it is fast, free and fully private. When you outgrow it, Forge and ComfyUI are waiting. Until then, test your prompts on our free browser generator and build your checkpoint shortlist from our best checkpoints and best LoRAs lists.
Frequently asked questions
Is Fooocus uncensored for NSFW?
Yes, when run locally. Fooocus has no hard NSFW safety filter blocking generation on your own GPU, so with a Pony, SDXL or Illustrious checkpoint it produces exactly what you prompt. There is no platform policy, credit meter or moderation. You only need to disable the default prompt expansion and styles, which interfere with anime models rather than censoring content.
Where do I put NSFW checkpoints in Fooocus?
Place checkpoint .safetensors files in the Fooocus\models\checkpoints folder, and LoRA files in Fooocus\models\loras. After adding them, click the refresh icon next to the model dropdown or restart run.bat, and your new checkpoints appear in the Base Model dropdown inside the Advanced panel. Select your Pony or Illustrious checkpoint there before generating for proper NSFW output.
How much VRAM does Fooocus need?
The official minimum is 4GB of Nvidia VRAM, with 6GB or more recommended for reasonable speed. Fooocus is efficient and runs on modest cards, but a 12GB RTX 3060 is the value sweet spot, giving fast Pony and Illustrious generation with headroom for LoRAs. On 4GB to 6GB cards, use the Extreme Speed mode and smaller resolutions, then upscale your keepers.
Can Fooocus run Pony and Illustrious models?
Yes. Drop the Pony Diffusion or Illustrious XL checkpoint into models\checkpoints, restart, and select it as the Base Model in the Advanced panel. Fooocus 2.5.0 and later added a dedicated Pony preset for sensible defaults. Crucially, disable prompt expansion and default styles, since those inject SDXL-style language that anime models trained on booru and score tags do not respond to well.
How do I stop Fooocus from changing my prompt?
Disable prompt expansion. By default Fooocus rewrites and expands prompts with extra tokens through its built-in GPT-style expansion, which helps casual SDXL prompts but hurts Pony and Illustrious. In the Advanced panel, turn off the Fooocus V2 expansion style and clear the default styles so your exact booru and score tags reach the model intact, giving clean, predictable anime results.
Why are my Fooocus anime images coming out blurry or generic?
Usually prompt expansion is still enabled or you are writing SDXL-style sentences instead of tags. Anime checkpoints like Pony and Illustrious expect precise booru and score tags. Disable expansion, clear default styles, confirm you selected the correct anime base model, and write proper tags. Also check you are not stuck on the default general model rather than your downloaded anime checkpoint.
Fooocus vs Forge vs ComfyUI, which should I use?
Start with Fooocus for the easiest uncensored local generation with smart defaults. Move to Forge, an optimized Automatic1111 fork, when you want full control with every sampler, ControlNet and extension while staying fast. Graduate to ComfyUI when you need custom node-based pipelines and advanced workflows. All three run the same checkpoints locally with no filter; they differ only in control and complexity.
How do I fix CUDA out of memory in Fooocus?
Your GPU VRAM is exhausted. Lower the resolution, set batch size to 1, switch to a faster performance mode like Extreme Speed, and close other GPU-heavy apps and browser tabs. Fooocus already offloads aggressively, so reducing resolution is the quickest fix. If it persists on a low-VRAM card, generate small and upscale the images you want to keep afterward.
