How to Set Up Stable Diffusion for Uncensored Output

By Alexandra Joly · Senior Editor · Updated May 17, 2026 · Reading time 11 min · Reviewed by editorial team

What does it take to set up Stable Diffusion for uncensored output?

This guide is for the reader who wants to run Stable Diffusion locally and skip every hosted middleman. The mainstream tools just say no: DALL-E declines sexual content, Midjourney's Community Guidelines prohibit it, Google Imagen documents its responsible-AI filters at the API layer. Run inference on your own GPU and you sidestep every guardrail at the platform layer. The only things left holding you back are the model card and the law. The cost is real too. Budget 30 to 90 minutes the first time you set it up, you need decent hardware (laptop-only readers are out), and you'll have to maintain the install when WebUI updates ship. If that's not your idea of a fun evening, skip straight to the four turnkey hosted alternatives at the bottom.

What you need before starting

Hardware floor. An NVIDIA GPU with at least 8 GB of dedicated VRAM is the entry ticket for SD 1.5 checkpoints (Realistic Vision V6.0 B1, DreamShaper 8). SDXL checkpoints (Pony Diffusion V6 XL, base SDXL, FLUX.1-dev) need 12 GB VRAM minimum, with 16-24 GB recommended for comfortable iteration [Source: NVIDIA: CUDA GPU Compute Capability · verified 2026-05-17]. AMD GPUs work via DirectML on Windows or ROCm on Linux, with reduced toolchain support and slower performance. Apple Silicon Macs run Stable Diffusion via Mochi Diffusion or the HuggingFace Diffusers library, but they lag NVIDIA on speed by roughly 2-3x at equivalent price points. CPU-only inference technically works, but the per-image time runs 10 to 50 times slower than a mid-range GPU. Don't bother for anything iterative.

Software floor. Python 3.10.6 is the version every official WebUI repo pins. Install something newer and you'll break dependencies and get cryptic import errors that eat a whole debug evening (I've done it, it's miserable). Git is required to clone the repos. A solid-state drive with at least 100 GB free is comfortable: a typical local install lands at 30-60 GB once you stack a few checkpoints and LoRAs.

Legal and ethical floor. Three universal red lines apply to every Stable Diffusion checkpoint, every LoRA, and every hosted alternative. First, no depictions of minors. US federal law under 18 USC 2252A treats AI-generated visual depictions of minors involving sexual activity as child sexual abuse material, with criminal penalties identical to photographic CSAM [Source: US DOJ: 18 USC 2252A statute · verified 2026-05-17]. Second, no non-consensual scenarios. Emerging state laws on non-consensual intimate imagery criminalize generation as much as distribution. Third, no real-person likeness without explicit consent. At least 27 US states now have laws covering non-consensual intimate imagery made with AI tools, and the federal Take It Down Act passed in 2025 layers a 48-hour takedown obligation on platforms. The FLUX.1-dev license spells out these out-of-scope uses directly in the official model card [Source: HuggingFace: black-forest-labs/FLUX.1-dev model card · verified 2026-05-17], and the Pony Diffusion V6 XL Modified Fair AI license adds commercial restrictions on top. None of this is a grey area. It applies whatever you generate, wherever you generate it.

Choose your WebUI

Three official WebUI repos serve the local install workflow, and which one you pick is honestly the biggest call you'll make here. Each ships a different balance between speed-to-first-output, how much VRAM it eats, and how far you can push the workflow later.

Automatic1111 (by the dev who goes by AUTOMATIC1111 on GitHub) is the historical reference, the WebUI most tutorials and YouTube walkthroughs target [Source: GitHub: AUTOMATIC1111/stable-diffusion-webui · verified 2026-05-17]. Forge by lllyasviel is a faster, lighter fork that keeps compatibility with most Automatic1111 extensions, and it's the pick I'd recommend when 8-12 GB VRAM is your budget [Source: GitHub: lllyasviel/stable-diffusion-webui-forge · verified 2026-05-17]. ComfyUI by comfyanonymous is a node-based interface. Steepest learning curve of the three, but the cleanest path to reproducible workflows once you're into inpainting chains, latent upscale, ControlNet stacks, and animation pipelines [Source: GitHub: comfyanonymous/ComfyUI · verified 2026-05-17].

For a first install, Automatic1111 or Forge wins on getting you to a first image fastest. Move to ComfyUI later, once you actually understand the diffusion pipeline and want control over every single node.

The 7-step install walkthrough

1. 1

   Verify your GPU and VRAM

   Open Windows Task Manager → Performance → GPU (or nvidia-smi from a terminal on Linux and Windows). Confirm the dedicated GPU memory matches the floor for your target architecture: 8 GB for SD 1.5 checkpoints, 12 GB for SDXL, 16-24 GB for comfortable FLUX iteration. If the system reports only an integrated GPU (Intel UHD, AMD Vega APU, Apple M-series before M3 Pro), the local install path will be impractical; skip to the hosted alternatives section. Mac users: Apple menu → About This Mac → More Info shows the GPU and unified memory; 16 GB unified memory is the comfortable floor for SDXL via Diffusers.

   Skip if you already know your hardware passes the floor.

2. 2

   Install Python 3.10.6 and Git

   Download Python 3.10.6 from python.org/downloads/release/python-3106 (NOT a newer version; the dependency graph breaks on Python 3.11+) [Source: Python.org: Python 3.10.6 release page · verified 2026-05-17]. On Windows, tick "Add Python to PATH" during installation. Install Git from git-scm.com/downloads. From a fresh terminal, verify with python --version (should return Python 3.10.6) and git --version (any 2.x is fine). If python returns a 3.11 or 3.12 version, your system PATH is pointing at the wrong install; fix that before proceeding.

   Skip if python --version already returns 3.10.6.

3. 3

   Clone the WebUI repo of your choice

   Pick one WebUI from the comparison table above. Open a terminal in the folder where you want the install (a path without spaces or special characters is safer). Run one of the three clone commands:

   • Automatic1111: git clone https://github.com/AUTOMATIC1111/stable-diffusion-webui.git
   • Forge: git clone https://github.com/lllyasviel/stable-diffusion-webui-forge.git
   • ComfyUI: git clone https://github.com/Comfy-Org/ComfyUI.git

   This pulls roughly 200 MB of repo code. Switch into the cloned folder.

   Skip if you already have a WebUI cloned and just need to update with git pull.

4. 4

   First boot and base model download

   On Windows, double-click webui-user.bat (Automatic1111 and Forge) or run_nvidia_gpu.bat (ComfyUI). On Mac and Linux, run bash webui.sh (Automatic1111 and Forge) or python main.py (ComfyUI) from the cloned folder. The first launch installs a Python virtual environment, pip-installs roughly 40 dependencies, and downloads the Stable Diffusion 1.5 base model (≈ 4 GB) from HuggingFace. Total first-boot time runs 5-15 minutes on a typical residential connection. The terminal will eventually print a URL like http://127.0.0.1:7860 (Automatic1111 and Forge) or http://127.0.0.1:8188 (ComfyUI). Open it in a browser. The base install is complete.

   Skip if the WebUI is already running at the printed URL.

5. 5

   Download an uncensored checkpoint from Civitai

   Civitai is the community marketplace where uncensored fine-tunes live. Visit civitai.com, sign in (free), toggle the content-rating filter to show explicit content, and pick a checkpoint matched to your style:

   • Pony Diffusion V6 XL for anime, cartoon, furry: 285.7 million downloads, SDXL architecture, score_9 quality-tagging prompting system documented by the author [Source: Civitai: Pony Diffusion V6 XL model card · verified 2026-05-17].
   • Realistic Vision V6.0 B1 for photorealistic SD 1.5: 729,019 downloads, CreativeML Open RAIL++-M license [Source: Civitai: Realistic Vision V6.0 B1 model card · verified 2026-05-17].
   • DreamShaper 8 for general-purpose versatility on SD 1.5: 1,022,411 downloads [Source: Civitai: DreamShaper model card · verified 2026-05-17].
   • FLUX.1-dev for prompt adherence (needs community LoRA for uncensored output) [Source: HuggingFace: black-forest-labs/FLUX.1-dev model card · verified 2026-05-17].

   Click the green Download button and save the .safetensors file. Move it to the models/Stable-diffusion folder of your WebUI install (or models/checkpoints for ComfyUI). Back in the WebUI browser tab, click the refresh icon next to the model dropdown and select the new checkpoint.

   Skip if you only need a SFW workflow; the base 1.5 model handles SFW prompts cleanly.

6. 6

   Install a LoRA for style or character control

   A LoRA (Low-Rank Adaptation) is a small fine-tune file, typically 50-300 MB, that biases the base checkpoint toward a specific style, character archetype, or output property, without retraining the multi-gigabyte base. Civitai's LoRA tab hosts the community catalog. Pick a LoRA matched to the base architecture you downloaded in step 5: an SD 1.5 LoRA pairs only with an SD 1.5 checkpoint; an SDXL LoRA pairs only with an SDXL checkpoint. Architecture mismatch produces garbled output, not an error message.

   Download the .safetensors file. Move it to models/Lora (Automatic1111 and Forge) or models/loras (ComfyUI). In your positive prompt, reference the LoRA with the syntax <lora:filename:0.7>, replacing filename with the exact filename without the .safetensors extension, and 0.7 with the strength. The typical band is 0.6-0.8; below 0.4 the effect is weak, above 1.0 it overpowers other prompt tokens.

   Skip if you want raw base-checkpoint output without style modification.

7. 7

   Disable the safety checker and run your first generation

   Automatic1111 ships a safety_checker that blurs flagged outputs. Edit webui-user.bat on Windows or webui-user.sh on Mac and Linux. Find the COMMANDLINE_ARGS= line and add --disable-safe-unpickle --no-half-vae. The --no-half-vae flag fixes a black-image bug on some NVIDIA cards. Save the file and restart the WebUI by closing the terminal and re-launching. Forge ships without the safety checker by default, so step 7 is unnecessary there. ComfyUI doesn't load the safety checker in default workflows either.

   Write a positive prompt (style descriptors, subject, quality tags) and a negative prompt (bad anatomy, blurry, deformed, low quality, text, watermark is a sensible default). Set sampler to DPM++ 2M Karras (SD 1.5) or Euler A (SDXL / Pony V6 XL), steps to 25-30, CFG to 6-8, resolution to 512×768 (SD 1.5) or 1024×1024 (SDXL). Click Generate. The first image typically lands in 5-15 seconds on a mid-range GPU. Iterate on prompt weight, sampler choice, seed lock, and LoRA strength until the output matches intent.

   Skip if your WebUI is Forge; the safety checker is not present by default.

Common pitfalls (and how to avoid them)

Pitfall 1: Installing Python 3.11 or 3.12 instead of 3.10.6

This is the one that catches everyone. Newer Python versions ship breaking changes in the dependency graph (torch, xformers, transformers), so the WebUI's launch.py fails with cryptic import errors on first boot. The fix: uninstall the wrong Python, install 3.10.6 from python.org, then verify with python --version from a fresh terminal. If you've got multiple Python versions sitting side-by-side on Windows, use the Python launcher (py -3.10) to force the right interpreter when you launch the WebUI manually.

Pitfall 2: Out-of-memory errors on SDXL or FLUX with 8 GB VRAM

SDXL checkpoints (Pony V6 XL, base SDXL) want 12 GB VRAM for comfortable generation at 1024×1024, so an 8 GB card sits below the floor. You've got three ways out. Switch to Forge, which lazy-loads the U-Net and cuts VRAM by roughly 25% on SDXL. Or generate at 768×768 then upscale. Or just fall back to an SD 1.5 checkpoint like Realistic Vision V6.0 B1 that fits cleanly in 8 GB. FLUX.1-dev at the full 12B-parameter scale is a non-starter below 16 GB VRAM; the FP8 or NF4 quantized versions on HuggingFace are the workaround for 8-12 GB cards.

Pitfall 3: LoRA and checkpoint architecture mismatch

Stack an SD 1.5 LoRA on an SDXL checkpoint (or the other way round) and you get garbled output, color artifacts, and "monster anatomy" with no error message. Just bad images. Civitai labels each LoRA's base model at the top of the model card, so read that label before you download. The filename sometimes hints at the architecture, but it lies often enough that you should trust the model card instead.

Pitfall 4: Downloading checkpoints from third-party mirrors

Civitai and HuggingFace are your two authoritative sources, full stop. Random third-party mirrors hosting .safetensors files have surfaced .pickle malware buried in model weights more than once across 2023-2024. Remember that --disable-safe-unpickle flag from step 7? It bypasses the WebUI's pickle scanner, so pairing it with a checkpoint from some sketchy mirror is a genuine way to get yourself owned. Stick to Civitai and HuggingFace, and check the creator's profile age and download count before you trust a checkpoint with no provenance.

Pitfall 5: Confusing local privacy with universal legality

Running Stable Diffusion locally keeps you out of reach of platform-level moderation (nothing gets uploaded), but it doesn't hand you legal immunity. The three red lines from the legal floor section apply exactly the same whether you generate on a hosted platform or on your own machine. The privacy is real. The legal cover on out-of-scope content is not. Simple test: if you wouldn't generate it on a hosted platform under your real name, don't generate it locally either.

Hard red lines that apply everywhere

Three constraints apply to every Stable Diffusion checkpoint, every LoRA, every hosted alternative, every jurisdiction we cover. No exceptions.

No depictions of minors. US federal law (18 USC 2252A) treats AI-generated visual depictions of minors involving sexual activity as child sexual abuse material, with criminal penalties identical to photographic CSAM. The Take It Down Act passed in 2025 adds federal-level platform takedown obligations on top. The FLUX.1-dev license writes the underage-content protection regime directly into the model card. And any checkpoint creator on Civitai who ships a model that reproduces this content easily gets pulled by Civitai's moderation team within days of detection.

No non-consensual scenarios. At least 27 US states have passed laws criminalizing non-consensual intimate imagery, covering generation, distribution, or both. The FLUX.1-dev license lists it in the out-of-scope section, and the Pony Diffusion V6 XL Modified Fair AI license carries its own restrictions on the same axis.

No real-person likeness without explicit consent. A LoRA trained to reproduce the likeness of a real celebrity, or your ex, or anyone else who didn't agree to it, fails the consent test every time. Civitai's terms of service banned this category in the 2024 policy update, and every hosted platform in this space forbids real-person deepfakes in its terms too. The legal exposure under emerging state laws is real, and it keeps growing.

Turnkey alternatives for non-technical readers

The local install above is free if you own the hardware, but the setup tax (30-90 minutes the first time, plus the upkeep every time a WebUI updates) and the GPU floor (12 GB VRAM for SDXL) price out a big slice of readers. So here are four hosted platforms that produce uncensored image generation on publicly accessible plans, with verifiable corporate operators and active affiliate programs in our approved set. We graded each one using our scoring page.

Promptchan's uncensored image gen is the image-gen flagship of this space. The underlying models include FLUX.1, Stable Diffusion 3, and Pony, the exact architectures listed above, just fronted by a hosted UI that swallows the setup tax for you. You get 30-50 daily gems on the free tier with watermarked output, and the Plus tier (around $11.99/month) removes the watermark. V5 Video Engine sits on Pro [Source: Apple App Store: Promptchan listing · verified 2026-05-17]. Pick this one if you want raw image-gen output without touching a single checkpoint.

Candy.ai, composite 8.4 is the companion-bundled option: image generation plus persistent chat plus voice plus a persona you customize, on a yearly plan that works out to roughly $3.99/month. That's the cheapest yearly price in our covered hosted set for a bundled image-gen experience. Pick this one when image gen is one part of a wider AI companion workflow rather than the whole point. And yes, it does girlfriend or boyfriend personas, whichever you're after.

DarLink's full breakdown is the narrative-anchored option. Swiss-domiciled, built around Scenario plus Media output with a three-tier Living Memory system, and it embeds 4K-claimed inline images inside an ongoing story arc. Pick this one if you want your uncensored output woven into a narrative rather than spat out as standalone prompts.

Joi's full review is the character-consistency option. Its Face-Sync V4 is the strongest documented identity-lock in our test set: the same character holds across re-rolls and across Dream Clips video frames at near-4K resolution. Premium yearly works out to about $4/month. Pick this one when keeping a single character recognizable across dozens of generations is the thing you care about most. One warning though: Joi is geo-blocked in the United Kingdom, so UK readers should grab one of the other three [Source: Joi: UK unavailable landing · verified 2026-05-17].

Frequently asked questions

When this guide does not apply

If you don't own an NVIDIA GPU with at least 8 GB VRAM, an AMD card you're willing to wrestle through reduced tooling, or Apple Silicon with 16 GB unified memory, the local install path just isn't practical. Pick a hosted alternative from the four-pick section above instead. And if you're here to generate real people without their consent, anything involving minors, or non-consensual scenarios, stop now. That's out of scope everywhere, on your own machine and on every hosted platform alike. This guide assumes adult readers generating fictional, AI-rendered, consensual-adult characters.

---

Last verified May 17, 2026 · See errata log for any post-publish corrections · Editor: Alexandra Joly · Methodology v1.0 · Editorial process · Affiliate disclosure

Skip the GPU setup, try Promptchan free →

Bundled companion plus image gen: try Candy.ai →

Narrative-anchored image gen: try DarLink Ai →

Face-lock character consistency: try Joi →

Related reading

• the AI image gen full ranking
• Best Stable Diffusion Models for Uncensored Output: 5 tested picks
• Best Uncensored AI Image Generators (full ranking)
• Civitai vs Promptchan: open-source vs hosted comparison
• AI Companion Scoring System v1.0 methodology
• the Promptchan 8-dim review
• the full Candy.ai writeup
• Joi.ai's writeup