ComfyUI NSFW Workflow Guide: 7-Step Uncensored Setup

By Alexandra Joly, Senior Editor · Last verified May 17, 2026 · Reviewed by editorial team · See our editorial process and errata log

What ComfyUI is, and why power users prefer it for uncensored output

This comfyui nsfw workflow guide assumes you already know the basics of Stable Diffusion uncensored output (Pony V6 XL for anime, Realistic Vision for photoreal, FLUX.1-dev for prompt adherence). If you haven't yet, go read our companion piece on the best Stable Diffusion models for uncensored output first. This page sits on top of that base knowledge and covers the ComfyUI-specific part.

Two real audiences land on a comfyui nsfw workflow guide. One is already running Automatic1111, hit a ceiling, and wants the node-graph control. The other read about ComfyUI on Reddit or Civitai and wants to skip Automatic1111 entirely. Both get served here. The seven steps below produce a working uncensored workflow on a fresh machine with zero prior Automatic1111 install.

What you need before you start

Three hard prerequisites and one soft one. Let me run them in order so you know if your setup is even in the game before you download anything.

Hardware floor first. NVIDIA GPU with 8 gigabytes of VRAM for SD 1.5 checkpoints (Realistic Vision V6.0 B1, DreamShaper 8). 10 to 12 gigabytes for SDXL checkpoints (Pony Diffusion V6 XL, base SDXL). 16 to 24 gigabytes for FLUX.1-dev. AMD GPUs work via DirectML or ROCm with reduced toolchain support, and Apple Silicon Macs run ComfyUI via MPS but lag NVIDIA on speed. CPU-only inference technically generates, but it runs ten to fifty times slower than a mid-range NVIDIA GPU, so it isn't practical for the kind of iterative prompting you'll actually do.

Then the disk budget. Twenty gigabytes free minimum to start. A single SDXL checkpoint runs six to seven gigabytes, and FLUX.1-dev is twelve to twenty-three gigabytes depending on precision. Add LoRAs, VAEs, ControlNets, and upscaler models and the install easily reaches one hundred gigabytes within a month of regular use. I learned that one the hard way.

Operating system next. Windows 10 or 11 (portable build, simplest path), Linux (manual install via Python venv), or macOS (manual install, slower inference on Apple Silicon). The Windows portable build is the path I'd recommend for a first install, because it bundles Python, PyTorch, and CUDA support and just runs.

And the soft one: patience. Two to four hours of learning patience for the first workflow. The node graph metaphor isn't intuitive coming from a tabbed UI. If you've used Blender shader nodes, Unreal Material Editor, or Houdini, the curve flattens fast. If not, expect at least one black-image generation in your first session (usually a VAE mismatch). It's a rite of passage.

The 7-step ComfyUI uncensored workflow

ComfyUI uncensored setup: 7 steps from fresh install to reusable workflow JSON

1. 1

   Install ComfyUI on your machine

   Download the ComfyUI portable build from the official comfyanonymous/ComfyUI GitHub release page for Windows, or clone the repository and run a manual Python install on Linux and macOS [Source: GitHub: ComfyUI release page · verified 2026-05-17]. The Windows portable build ships a bundled Python 3.11 and PyTorch with CUDA support and runs out of the box on NVIDIA GPUs. Extract the archive, run run_nvidia_gpu.bat, and confirm the web UI loads in your browser on http://127.0.0.1:8188.

   Linux and macOS users install via the documented manual path: git clone https://github.com/Comfy-Org/ComfyUI, create a Python virtual environment, install PyTorch matched to your CUDA or ROCm or MPS backend, install the ComfyUI requirements, then run python main.py. Give it ten to fifteen minutes for the initial dependency install on a fresh machine.

   Skip this step only if you already have a working ComfyUI install on the machine you plan to run uncensored workflows on.

2. 2

   Install ComfyUI Manager for one-click node and model management

   Clone the ltdrdata/ComfyUI-Manager repository into the ComfyUI/custom_nodes/ folder, then restart ComfyUI [Source: GitHub: ltdrdata/ComfyUI-Manager · verified 2026-05-17]. After the restart, a new Manager button shows up in the ComfyUI menu. The Manager panel surfaces three things you'll lean on constantly: install custom nodes from the registry, install missing nodes from an imported workflow, and download checkpoints, LoRAs, and VAEs straight from Civitai or HuggingFace.

   Without ComfyUI Manager, every custom node and missing dependency turns into a manual file-copy chore. And every published Civitai workflow you import will reference at least one custom node, so the Manager's automatic missing-node detection saves you hours per workflow. Don't skip it.

   Skip this step only if you have a hard policy against installing community-maintained third-party tools. ComfyUI Manager is open source, widely audited, and the de-facto standard add-on for ComfyUI.

3. 3

   Download Pony Diffusion V6 XL or Realistic Vision V6 to the right folder

   For anime, cartoon, or furry uncensored output, download Pony Diffusion V6 XL from Civitai and place the safetensors file in ComfyUI/models/checkpoints/ [Source: Civitai: Pony Diffusion V6 XL model card · verified 2026-05-17]. It trained on roughly 2.6 million images with a 1:1 ratio between anime, cartoon, furry, and pony datasets and a 1:1 ratio between safe, questionable, and explicit ratings. The training set is half explicit by design, which is why the checkpoint produces uncensored output without any LoRA stacking.

   For photorealistic uncensored output on lower VRAM hardware, download Realistic Vision V6.0 B1 instead [Source: Civitai: Realistic Vision V6.0 B1 model card · verified 2026-05-17]. The 729,019-download SD 1.5 checkpoint by SG_161222 is the go-to photoreal pick at the 8 gigabyte VRAM level, and it pairs well with Face Detailer in step 5.

   Optional: if you've got 16-plus gigabytes of VRAM and you care more about prompt adherence than a specific style, download FLUX.1-dev from HuggingFace and an uncensored community LoRA from Civitai [Source: HuggingFace: black-forest-labs/FLUX.1-dev model card · verified 2026-05-17]. FLUX base is SFW-trained, so the uncensored output comes from the LoRA.

   Skip this step only if you already have a Stable Diffusion checkpoint in ComfyUI/models/checkpoints/. Mixing model architectures (SDXL vs SD 1.5 vs FLUX) in the same workflow isn't supported by the base ComfyUI graph, so pick one architecture per workflow.

4. 4

   Build the base text-to-image graph in ComfyUI

   Drop the following nodes onto the canvas in this order: Load Checkpoint (select your downloaded Pony V6 XL or Realistic Vision file), CLIP Text Encode (Positive Prompt), CLIP Text Encode (Negative Prompt), Empty Latent Image (1024 by 1024 for SDXL, 512 by 768 for SD 1.5), KSampler (steps 25 to 30, CFG 6 to 8 for Pony V6 XL or 1.5 to 2.0 for Realistic Vision with the Hyper LoRA), VAE Decode, and Save Image.

   Wire them in the standard chain: Load Checkpoint sends MODEL to KSampler, CLIP to both prompt encoders; positive prompt sends CONDITIONING to KSampler positive input; negative prompt sends CONDITIONING to KSampler negative input; Empty Latent Image sends LATENT to KSampler latent input; KSampler sends LATENT to VAE Decode; VAE Decode sends IMAGE to Save Image.

   For Pony V6 XL specifically, set the positive prompt to begin with the score tag sequence the model author documented: score_9, score_8_up, score_7_up, source_anime, [your prompt here]. Those score tags are a quality-modifier convention the community standardised on, and if you omit them your output comes out visibly degraded. Use the Euler A sampler at 25 steps and CFG 7 for the strongest first pass.

   Skip this step's wiring detail only if you imported a working workflow JSON from Civitai. In that case the graph is already built and you jump straight to step 5.

5. 5

   Add a Face Detailer node for face fidelity

   Faces shrink to a small fraction of the canvas at 1024 resolution and lose fidelity on the first pass. Install Impact Pack via ComfyUI Manager to get FaceDetailer, the ComfyUI equivalent of ADetailer for Automatic1111. ADetailer is the canonical face-detail tool in the Automatic1111 world, and FaceDetailer mirrors its behaviour inside the ComfyUI node graph [Source: GitHub: Bing-su/adetailer (Automatic1111 reference) · verified 2026-05-17].

   Wire FaceDetailer after VAE Decode and before Save Image. The node crops detected faces using a YOLO face-detection model, re-samples them at higher resolution through the base checkpoint, then composites the upscaled face back into the original frame. Default settings (denoise 0.4, steps 20, CFG 7) work well for most uncensored portraits. Raise denoise to 0.5 if faces still look soft, or drop it to 0.3 if the redraw drifts too far from the original.

   For workflows that generate several characters per frame, FaceDetailer detects and processes each face on its own, which is exactly what you want for the multi-character compositions that set Pony V6 XL apart from single-subject SDXL.

   Skip this step only if your prompts target full-body shots where faces are intentionally small or absent. For any portrait, half-body, or close-up workflow, FaceDetailer is the highest-payoff second pass you can add.

6. 6

   Add a 2x upscale node for production-quality output

   Stable Diffusion XL outputs at 1024 pixel native and SD 1.5 at 512 pixel native, and plenty of production work needs higher than that. Install Ultimate SD Upscale via ComfyUI Manager and add it after FaceDetailer. It tiles the image into overlapping segments, re-samples each tile through the base checkpoint at a lower denoise strength (0.2 to 0.35), then stitches the tiles back together with seamless blending.

   The output comes out cleaner than naive bilinear or Lanczos upscaling, and it keeps the texture detail that simpler upscalers blur away. For SDXL workflows, 2x upscaling lifts 1024 to 2048 pixel; for SD 1.5 workflows, 2x lifts 512 to 1024 or 768 to 1536. VRAM pressure spikes during the upscale pass, so if you're on an 8 to 10 gigabyte card, expect to use the --medvram flag at ComfyUI startup or accept a slower tile-by-tile pass.

   If you'd rather use a pre-trained model approach, 4x-UltraSharp and RealESRGAN_x4plus_anime_6B are both available through ComfyUI Manager's model browser. They run faster than Ultimate SD Upscale, though they apply more aggressive sharpening, which some workflows love and others can't stand.

   Skip this step only if you don't need output above native checkpoint resolution. For social-media-sized output (1024 by 1024), the FaceDetailer pass on its own is plenty.

7. 7

   Save the workflow JSON and reuse it across sessions

   Click Save in the ComfyUI menu to export the entire node graph as a workflow JSON file. The JSON captures every node, every connection, every parameter value, and every model reference in a single file. You can re-import it on the same machine by dragging the file onto the canvas, share it with collaborators who run their own ComfyUI install, or attach it to generated images so anyone with the same models can reproduce the exact graph [Source: ComfyUI documentation · verified 2026-05-17].

   This is honestly the most underrated ComfyUI feature. Every published Civitai uncensored workflow ships as a downloadable JSON you drop straight onto the canvas. The Civitai community workflow gallery hosts thousands of these files covering Pony V6 XL, FLUX, and Realistic Vision setups with custom LoRAs, ControlNets, and post-processing chains [Source: Civitai: community Stable Diffusion content · verified 2026-05-17].

   Two production habits the JSON workflow unlocks. First, version your workflows in a git repository so you can roll back to a prior graph when a new node breaks a setup that was working fine. Second, build a small library of three to five JSON files (one for portraits, one for full-body, one for upscale-only, one for inpainting, one for FLUX) and just load the right one per session instead of rebuilding the graph from scratch every time.

   Skip this step only if you plan to use ComfyUI as a one-off image generator and never reproduce a workflow. For any iterative or production use, JSON export is non-negotiable.

Common pitfalls (and how to avoid them)

Pitfall 1: wrong VAE for the checkpoint

The most common first-session failure is a black image, and the cause is almost always a VAE mismatch. SDXL checkpoints expect the SDXL VAE, SD 1.5 checkpoints expect the SD 1.5 VAE, and FLUX expects a FLUX VAE. ComfyUI loads a default VAE with the checkpoint, but some Civitai downloads ship without the embedded VAE and hand you a black output until you explicitly wire a Load VAE node into the graph. So when a workflow spits out pure black or pure noise on the first generation, check the VAE wiring before you touch the prompt or the sampler.

Pitfall 2: skipping the score tags on Pony V6 XL

Pony Diffusion V6 XL was trained with a quality-tagging system the model author documented (score_9, score_8_up, score_7_up, score_6_up). Prompts that leave the score tags off come back visibly degraded: muddy colours, soft anatomy, off-style results that people blame on the model when it's really a prompting convention they skipped. Every Pony V6 XL positive prompt should begin with the score tag sequence, and the negative prompt usually includes score_4, score_3, score_2, score_1 to push the sampler away from low-quality output.

Pitfall 3: running SDXL on 8 GB VRAM without the --medvram flag

SDXL checkpoints can technically run on 8 gigabytes of VRAM, but the default ComfyUI startup tries to load the full model into VRAM and runs out mid-generation. The fix is the --medvram flag at startup (on Windows, edit run_nvidia_gpu.bat to append the flag; on Linux, append it to the python main.py command). It swaps the model between VRAM and RAM at a small speed cost. If you hit a CUDA out-of-memory error in the console mid-generation, reach for --medvram before you downgrade to a smaller checkpoint.

Pitfall 4: installing custom nodes without ComfyUI Manager

Here's a story I've watched play out a dozen times. Someone lands on a Reddit thread, sees a recommended custom node, clones it straight into ComfyUI/custom_nodes/, and the next restart fails to load because of a dependency conflict with another node. ComfyUI Manager handles dependency resolution and version pinning for you; a manual git clone doesn't. Run every custom-node install through the Manager's Install Custom Nodes panel, and save the manual git clone for nodes that aren't in the registry yet.

Pitfall 5: treating workflow JSON files as untrusted code

A workflow JSON imported from an unknown Civitai user can reference custom nodes that run arbitrary code on your machine when the workflow runs. The JSON itself is just data and can't execute anything; the referenced custom nodes are Python files and absolutely can. Import workflow JSONs from named Civitai authors with a verified upload history, and let ComfyUI Manager install the referenced nodes from the registry instead of cloning from random GitHub URLs.

When this guide doesn't apply

This is the technical setup path for power users, and it doesn't apply to everyone. Be honest with yourself about the following.

You don't have an NVIDIA GPU with at least 8 gigabytes of VRAM. The workflow above needs CUDA hardware for usable generation speeds. AMD and Apple Silicon paths exist, but with reduced toolchain support and slower iteration.

You don't have two to four hours to spend on the first workflow. ComfyUI isn't a fifteen-minute install. A realistic time-to-first-generation on a fresh machine runs an hour for the install and dependencies, an hour for the first base workflow, an hour to add FaceDetailer and upscale, plus at least thirty minutes to troubleshoot the inevitable VAE or VRAM error.

You want a workflow you don't have to maintain. Custom nodes update, base checkpoints update, ComfyUI itself updates, and a JSON workflow that worked fine in winter can break by spring. Hosted alternatives (Promptchan, Candy.ai, DarLink Ai, Joi) carry that maintenance burden on their side.

You generate the occasional image rather than iterating in batches. The setup tax only pays off across regular use. For one image a month, the four turnkey alternatives at the footer produce equivalent uncensored output without the GPU and without the install.

If any of that is you, the four hosted picks at the footer are the right pivot. The seven-step workflow above is for the readers who clear all four conditions.

ComfyUI vs Automatic1111: which one should you pick?

Go with ComfyUI when you need precise control over a multi-step pipeline. Think text-to-image into face detailer into upscale into post-process, or text-to-image into ControlNet pose conditioning into a LoRA stack into custom sampler scheduling. Each operation is its own node you can re-route, duplicate, or disable, and the whole graph exports as a JSON file that another ComfyUI user reproduces identically.

Go with Automatic1111 when you want to load a checkpoint, type a prompt, click Generate, and see the result. The tabbed UI hides the node graph behind a simpler interface. ADetailer, upscale, and other extensions install through the Extensions tab and show up as additional tabs rather than nodes. Shallower learning curve, lower customisation ceiling.

Plenty of people run both: ComfyUI for the iterative pipeline work, Automatic1111 for quick one-offs. Point them at the same disk location and they share the same checkpoint, LoRA, and VAE folders, so you're not duplicating model files all over your drive.

There's a third option some folks land on too. Forge is a community fork of Automatic1111 tuned for lower VRAM and faster sampling. It runs the same checkpoints and the same ADetailer extension, and it's worth a look if the standard Automatic1111 install path has been giving you grief.

Frequently asked questions

If you don't want to manage a ComfyUI install: four turnkey alternatives

The seven-step workflow above is genuine power-user territory. If you landed here without an NVIDIA GPU, without two to four hours of learning patience, or without any appetite for babysitting checkpoint updates and custom-node dependency conflicts, you've got a legitimate alternative path. Hosted platforms produce uncensored image output on a publicly accessible plan, with a real company behind them and an active affiliate offer. The four picks below come straight from our scoring.

1. Promptchan's uncensored image gen: image-gen specialist (Approved PPS)

Recommended (7.0-7.9) under our AI Companion scoring. Top quartile on Image Generation Quality. The underlying models include Flux.1, Stable Diffusion 3, and Pony, the same architectures from the seven-step workflow above, just fronted by a hosted UI that strips away the setup tax. The 20-million-plus user-generated public gallery is a genuine draw, because you get to browse community prompts before paying a cent. You get 30 to 50 daily gems on the free tier with watermarked output, and the Pro tier runs ≈ $26.99 per month [Source: Apple App Store: Promptchan listing · verified 2026-05-17].

The catch: the mainstream chat is bolted on, and third-party reviewers call it short and generic. Promptchan is an uncensored image-and-video tool with a chat veneer, not a chat-first AI companion.

Try Promptchan free →

2. Candy.ai review: companion app with bundled image gen

Excellent (8.0-8.9) under our AI Companion scoring. Top quartile on Image Generation Quality and Customization. The yearly plan lands at ≈ $3.99 per month effective, the lowest of any hosted option we cover for a companion plus image-gen bundle. The UX is polished, with visual customization sliders, scene-change prompts, and re-roll consistency that all add up to a strong first impression. And the bundled experience (companion creation, image gen, voice) is exactly what most people are after when they search "AI girlfriend with image generation."

The catch: the memory weakness shows up across every hands-on reviewer, so expect the persona to drift past week 2 to 3 of heavy use.

Try Candy.ai →

3. our DarLink pick: narrative-anchored scenario + media

Recommended (7.0-7.9). Strong on Customization and Image Generation, mid-pack on Voice. It's a Swiss-domiciled, narrative-first platform built around Scenario plus Media output. The three-tier Living Memory system holds onto the persona better than the impulse-buy hosted alternatives do. The inline 4K-claimed image generation is the standout here, and the whole thing is aimed at readers who want their uncensored images woven into an ongoing story rather than dropped out as standalone prompt results.

The catch: it's less of a fit for the raw prompt-and-output grind where Promptchan rules. The narrative wrapper is the point, not the friction-removal.

Try DarLink Ai →

4. Joi.ai review: face-lock + Dream Clips

Excellent (8.0-8.9). Top quartile on Image Generation Quality and Video Generation. Face-Sync V4 is the strongest documented identity-lock in our test set, holding the same character across re-rolls and across Dream Clips video frames at near-4K resolution. If your main concern is character consistency (the same problem ComfyUI users wrestle with using face-embedding LoRAs and Face Detailer), Joi is the managed-platform answer. The free tier gives you ≈ 5 messages a day and 6 image generations with no explicit output, and the Premium yearly plan runs ≈ $4 per month effective.

The catch: the Neuron token economy stacks up fast, and live video calls run roughly $30 to $50 per hour. UK readers are geo-blocked, with Joi serving a /uk/unavailable landing instead of implementing the UK Online Safety Act [Source: Joi: UK unavailable landing · verified 2026-05-17].

Try Joi →

---

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 ComfyUI setup, try Promptchan free →

Related reading

• AI image generation hub and full ranking
• Best Stable Diffusion models for uncensored output (5 tested picks)
• Top 10 uncensored AI image generators
• How we score AI companion apps (our methodology)
• Promptchan full review (Pro tier and Video Engine)
• Candy.ai
• Joi
• DarLink AI review