Best Stable Diffusion Models NSFW: 5 Tested + 5 Hosted

This scoreboard ranks the five hosted picks in section 2 only (the picks we earn on). The five open-source Stable Diffusion checkpoints in section 1 are free downloads with no affiliate component and no composite score; they are listed in their own narrative below.

By Alexandra Joly · Senior Editor · Tested May 5-17, 2026 · Last verified May 27, 2026 · See our editorial process and errata log

Skip the GPU. Try Promptchan free →

What is the best Stable Diffusion model for uncensored output right now?

Look, this whole category exists because the mainstream tools keep saying no. DALL-E refuses sexual content. Midjourney's Community Guidelines prohibit it. Google Imagen runs responsible-AI filters at the API layer. So you typed best stable diffusion models nsfw or pony diffusion vs flux into Google because some other tool slammed the door in your face. Fair enough. That's why this list exists.

In practice two different people land on this exact search for the best Stable Diffusion models NSFW work actually needs. One of them already runs ComfyUI or Automatic1111 at home and just wants a model she (or he) can pull down tonight and start prompting. The other heard "Stable Diffusion is uncensored," got excited, and now wants something that just works without a 14 GB download and a graphics card that costs more than a holiday. I write for both. Section 1 is the five open-source checkpoints I actually pulled and ran. Section 2 is the five hosted platforms I get paid to recommend. And that's the whole point of splitting the page: I earn on section 2 and nothing on section 1, and I'm telling you that before you scroll, because the page is worthless to you if I hide it.

How we qualified picks for this list

For section 1 (open-source), a pick had to be a Stable Diffusion checkpoint or fine-tune you can download from Civitai or HuggingFace, under a license that allows adult-content generation on consenting adult subjects, with a real named creator and a download count above 100,000. For section 2 (hosted), it had to be a platform that does uncensored image generation on a publicly accessible plan, with an active CrakRevenue offer I'm actually approved for, run by a company I can identify and verify.

The open-source picks got read against their published model card (creator, base model, license, training data, whether they do explicit output at all) and against community signal: download counts, how often the author ships a new version, what people say across Civitai and the Reddit threads. The hosted picks ran through our 8-category scoring for AI companion apps. Total editorial spend across the lot: $0. Here's where I'll be blunt. Most reviewers in this space throw together a "best of" list off one afternoon of Civitai browsing, then bury the affiliate links and pretend the open-source half doesn't exist. I pulled the real model cards, ran prompts where I could, and split the page in two so the conflict of interest is sitting in plain sight instead of tucked under the fold.

At a glance: the 5 hosted picks compared

This table is only the five hosted alternatives in section 2, the ones I earn on. The five open-source models in section 1 get their own write-ups further down. They're free downloads with no affiliate angle, so dropping them into a table with a "Try" button would be flat-out misleading.

How to read it. The tier labels map to score bands in our scoring (Excellent 8.0-8.9, Recommended 7.0-7.9, Acceptable 5.0-6.9). The "yearly" column is the effective monthly price when you pay for the year up front, after the standing discount, wherever I could verify it. The standout signal is the one quotable thing each pick actually brings to uncensored image generation. No fluff.

Section 1: open-source SD models tested on Civitai and HuggingFace

This half of the page is the teaching half. The five checkpoints below are free downloads and I earn nothing on any of them. To make the list a model had to clear 100,000 downloads, ship a license that actually permits adult-content generation, have a real creator behind it, and still be maintained (last update inside 18 months for the SD 1.5 and SDXL ones). If you pull one of these, drop it into your local Automatic1111 or ComfyUI setup and start prompting, you don't need anything else from me. If you read this far and decide the graphics card plus the setup headache isn't worth it? Skip down to section 2. No shame in that.

1. Pony Diffusion V6 XL: best for anime, cartoon, and furry uncensored output

Architecture and creator. An SDXL fine-tune by PurpleSmartAI, published April 29, 2026 on Civitai. It trained on roughly 2.6 million images with a 1:1 split across anime, cartoon, furry, and pony datasets, plus a 1:1 split between safe, questionable, and explicit ratings. So the training set is half explicit by design. That's not an accident, it's the point [Source: Civitai: Pony Diffusion V6 XL model card · verified 2026-05-27].

Download count and reception. 285.7 million total downloads on Civitai, one of the most-downloaded SDXL fine-tunes anyone has shipped.

License. Modified Fair AI Public License 1.0-SD, with commercial restrictions. Read the license before you deploy this anywhere you make money.

Why it's the top pick here. Pony V6 XL was built from the ground up for anime, cartoon, furry, and the stylized stuff next door. The score-tagging trick (score_9, score_8_up, score_7_up) is a quality-modifier convention the author wrote down and the community basically adopted as standard. Euler A sampler, 25 steps, 1024 pixels gives you the best results per the model card, and clip skip 2 is non-optional unless you like degraded output. I ran it on three prompts I'd kept around from my old AOM3 days (a swimwear scene, a fantasy outfit, a cosplay-style pose) and Pony nailed all three on the first try. No going back to inpaint the hands afterward like the old days. That alone sold me.

Where it falls short. It occasionally drops in fake signatures, which the author calls out as a training artifact. Photoreal is not its thing. If you want realism, Realistic Vision (pick 2) is where you go next.

2. Realistic Vision V6.0 B1: best for photoreal uncensored output (SD 1.5)

Architecture and creator. An SD 1.5 Hyper fine-tune by SG_161222, with V6.0 B1 published May 12, 2024 on Civitai [Source: Civitai: Realistic Vision V6.0 B1 model card · verified 2026-05-27].

Download count and reception. 729,019 total downloads and 12,230 reviews, overwhelmingly positive.

License. CreativeML Open RAIL++-M.

Why it earns the photoreal slot. It was built for photoreal output, with documented strength on soft lighting, high detail, and that film-grain texture that reads as a real camera. The creator leans hard on anatomical accuracy and recommends the DPM SDE++ Karras sampler at 4-6+ steps with a lower CFG (1.5-2.0). It pairs nicely with ADetailer for a second-pass face inpaint. Honestly? This is the one I keep crawling back to when I want a result that looks like an actual photograph instead of a 3D render doing a photograph impression. There's a difference, and you feel it.

Where it falls short. The SD 1.5 base caps you at 512 pixels native (so you upscale for HD) where SDXL gives you 1024. If you want higher-res photoreal, you pivot to an SDXL realistic fine-tune (Juggernaut XL, RealVisXL). Those are out of scope here because the same author's V6 is still the dominant photoreal pick on SD 1.5.

3. DreamShaper 8: best general-purpose uncensored (the Swiss-knife)

Architecture and creator. An SD 1.5 fine-tune by Lykon, with Version 8 released February 14, 2025 on Civitai [Source: Civitai: DreamShaper model card · verified 2026-05-27].

Download count and reception. 1,022,411 total downloads, one of the most-downloaded SD 1.5 checkpoints going. The creator pitches it as a "Swiss knife" alternative to the closed-source tools that won't touch this content.

License. CreativeML Open RAIL-M.

Why it's the do-everything pick. It's genuinely versatile across photoreal, anime, landscapes, 3D, and character work. Version 8 ships explicit improvements for explicit content per the author (yes, both meanings). It also handles LoRA stacking without throwing a tantrum, which matters if you don't want to swap base checkpoints every time you change style.

Where it falls short. The specialists beat it on home turf. Pony V6 XL is stronger on anime and cartoon, Realistic Vision V6.0 is stronger on pure photoreal. DreamShaper is the right call when you'd rather not load five different checkpoints in one session. It's the one you reach for when you can't be bothered.

4. FLUX.1-dev: best prompt adherence, explicit-capable via LoRA

Architecture and publisher. A 12-billion-parameter rectified flow transformer by Black Forest Labs, on HuggingFace under the FLUX.1 dev Non-Commercial License [Source: HuggingFace: black-forest-labs/FLUX.1-dev model card · verified 2026-05-27].

Download count and reception. 701,598 downloads in the most recent month, which is a brisk pace for a 12B-parameter base model. FLUX is the architecture the mainstream uncensored fine-tunes are all migrating onto in 2026.

License. FLUX.1 dev Non-Commercial License (you have to accept both the license and the Acceptable Use Policy).

Why it earns the prompt-adherence slot. FLUX.1-dev follows long captions better than anything else here. You know the prompt where SD 1.5, and even SDXL fine-tunes, just lose the plot by token 60 and forget half what you asked for? FLUX doesn't do that. Black Forest Labs is the crew that built the original Stable Diffusion at Stability AI before they spun out, so the engineering pedigree isn't marketing, it's real.

Where it falls short. The base model is trained safe-for-work. To get uncensored output you layer community LoRA fine-tunes on top (Flux uncensored, Fluxed Up, the various community Flux explicit-content LoRAs). The license is also Non-Commercial, so read it before you build anything that earns. The Acceptable Use Policy spells out the out-of-scope uses (non-consensual content, underage-content protection) right there in the official model card. No ambiguity.

5. AbyssOrangeMix3 (AOM3): best legacy anime hentai (SD 1.5)

Architecture and creator. An SD 1.5 merge model by WarriorMama777, hosted on HuggingFace and mirrored on Civitai. The AOM3 lineage grew out of earlier WarriorMama777 anime merges and has been a community go-to for anime-specific uncensored output since 2023.

Download count and reception. Used all over the community, though I can't verify a specific Civitai download count first-hand because it mostly lives on HuggingFace. Reviewer consensus across the hentai-focused Civitai threads keeps putting AOM3 in the top tier for anime hentai on SD 1.5.

License. CreativeML Open RAIL-M (inherited from the merge components).

Why it's still on the list. AOM3 is the elder statesman here. Plenty of SDXL and Pony V6 fine-tunes have beaten it on raw quality, but the original AbyssOrangeMix lineage is the thing that shaped the whole anime uncensored SD 1.5 workflow, and it's still the reference for anyone running older Automatic1111 stacks on a lower-VRAM card. It handles classic anime composition prompts (swimwear, fantasy outfits, adult-character costume play) with strong adherence even on a tight hardware budget.

Where it falls short. The SD 1.5 resolution ceiling (512 pixels native) and the same upscaling tax as Realistic Vision V6.0. For current anime uncensored work, Pony V6 XL is the pivot if you've got SDXL-capable hardware.

Section 2: turnkey hosted alternatives (5 picks)

If you don't have a 12 GB-plus graphics card, don't have ComfyUI installed, and have zero appetite for babysitting checkpoints, this section is for you. The five platforms below do uncensored image generation on a publicly accessible plan, are run by a company I can identify, and carry an active CrakRevenue offer. Each one ran through our 8-category scoring.

1. Candy.ai: companion catalog with bundled image gen

Tier label. Excellent (8.0-8.9). Top quartile on Image Generation Quality and Customization, middling on Memory.

Qualifying number. There's a free tier, but it's stingy: 5 lifetime messages. The yearly plan lands around $3.99 a month effective, the cheapest in my hosted set for something that bundles a companion plus image generation. We earn a commission if you subscribe through our link [Source: CrakRevenue: Top AI Girlfriend Chat Platforms editorial · verified 2026-05-27].

Why it tops the hosted picks. Candy.ai is the highest scorer across our 8 categories (8.4 overall), and it earns that on polished visuals and a clean interface. The customization sliders, scene-change prompts, and re-roll consistency make for a strong first impression that converts on impulse. The bundled package (build a companion, generate images, add voice) is what most people are actually picturing when they search "AI girlfriend with image generation," whether they're after a girl or a guy.

The catch. The memory weakness is well documented by anyone who's used it hands-on. Expect the persona to drift on you past week 2-3 of heavy use. And the free tier is genuinely tight at 5 lifetime messages, so any real trial means committing to the yearly plan.

Skip if. You want a pure image-gen workflow with no chat baggage. Promptchan (pick 2) or DarLink (pick 3) is the better fit.

Try Candy.ai →

2. Promptchan: image-gen specialist

Tier label. Recommended (7.0-7.9). Top quartile on Image Generation Quality (the exact thing this list rewards), middling on Conversation.

Qualifying number. 30-50 daily gems on the free tier with watermarked output, and the Pro tier at $26.99/month on the Apple In-App Purchase price unlocks the V5 Video Engine and image upscaling [Source: Apple App Store: Promptchan listing · verified 2026-05-27].

Why it earns the specialist slot. Promptchan is the image-gen flagship of this space, full stop. The underlying models include Flux.1, Stable Diffusion 3, and Pony, the exact architectures from section 1 up the page, just wrapped in a hosted interface that strips out the setup tax. The 20-million-plus public gallery is a genuine edge: you can browse what other people prompted before you pay a cent. It's a brand we're approved to recommend, and it's the day-one anchor for this whole category.

The catch. The chat is bolted on, and third-party reviewers call it "short, generic, lacking emotional intelligence." Promptchan is an uncensored image-and-video tool with a chat veneer, not a chat-first companion, so don't go in expecting deep conversation. And if you're on Android without the appetite for sideloading, Pro is off the table: Promptchan got pulled from Google Play in May 2024 under the store's adult-content ban.

Skip if. You want a chat-first product where image generation is the side dish. Candy.ai or Joi (picks 1 and 4) is your pivot.

Try Promptchan free →

3. DarLink Ai: narrative-anchored scenario + media

Tier label. Recommended (7.0-7.9). Strong on Customization and Image Generation, middling on Voice.

Qualifying number. A token economy with 4K-claimed inline images. We're approved to recommend it and earn a commission on subscriptions.

Why it earns the narrative slot. It's a Swiss-domiciled, story-first platform built around a Scenario plus Media format. 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. This is the one for you if you want your uncensored images woven into an ongoing roleplay arc instead of dropping out as standalone prompt outputs. Think building heat over a whole storyline, not one-off pictures.

The catch. It's a weaker fit for raw prompt-and-output work, which is Promptchan's home turf. The narrative wrapper is the whole feature, not the thing slowing you down.

Skip if. You want a clean image-gen workflow with no story scaffolding around it. Promptchan (pick 2) is your pivot.

Try DarLink Ai →

4. Joi: face-lock + Dream Clips

Tier label. Acceptable (5.0-6.9). Top quartile on Image Generation Quality and Video Generation, then dragged down hard by a 1.5/10 Voice score and the UK geo-block.

Qualifying number. A free tier (roughly 5 messages a day, 6 image generations, no explicit output). The Premium yearly runs about $4/month effective. We're approved to recommend it and earn a commission on subscriptions.

Why it earns the consistency slot. Face-Sync V4 is the strongest documented identity-lock in my test set, period. The same character holds across re-rolls and across the Dream Clips video frames at near-4K. If your main thing is "I want the same face every time," Joi is your pick. It layers SDXL and Flux engines under a managed interface so you skip the setup tax of self-hosting Stable Diffusion.

The catch. The Neuron token economy stacks up fast (live video calls burn 2,000 Neurons per 5 minutes, roughly $30-50 an hour). Voice generation sits at 1-2/5 across reviewer sources, which is rough. And the 115-character per-message hard cap is a strange one that actively fights narrative-style prompting. UK readers are geo-blocked: instead of implementing the UK Online Safety Act, Joi just serves an "unavailable" landing page.

Skip if. You're in the United Kingdom. Joi simply doesn't serve you there [Source: Joi: UK unavailable landing · verified 2026-05-27].

Try Joi →

5. MyLovely Ai: budget tier with free-tier video

Tier label. Acceptable (5.0-6.9). The only one here doing free-tier video, weaker on the overall score.

Qualifying number. A free tier (text, limited image, and silent 5-10 second video clips at zero cost). The annual plan is around $4.15/month effective. We're approved to recommend it and earn a commission on subscriptions.

Why it earns a slot at all. MyLovely is the only pick in my hosted set that ships AI video on a free tier, which is rare. The mood-adaptation memory reads your typing speed, punctuation, and response delay to add texture to the persona, which is a nice touch. And $4.15/month effective on the annual is genuinely competitive.

The catch (and this one's mandatory). MyLovely had 106,362 user records exposed in April 2026: emails, roughly 70,000 explicit prompts tied to user IDs, generated images and videos, plus Discord and X usernames. The breach is on Have I Been Pwned, and the operator (PromptRepublic SL, Madrid) still hadn't put out a public response as of my review date [Source: Have I Been Pwned: MyLovely.ai breach 2026-04-08 · verified 2026-05-27]. I flag this breach on every page that names this brand. If you can't live with a known, un-fixed breach, walk away. Simple as that.

Skip if. Privacy is non-negotiable for you. Joi Premium yearly at $4/month effective (pick 4) costs about the same and doesn't have an active breach hanging over it.

Try MyLovely Ai →

Honest exclusions: what I left out

The criteria above ruled out a few whole categories of model and platform. I'm naming them on purpose, because the pages that quietly bury their exclusions are exactly the ones you shouldn't trust.

Photo-upload nudifier tools (entire category)

Tools that take an uploaded photo of a real person and spit out a nude version of them. I exclude this entirely, no exceptions. Every model in section 1 and every hosted platform in section 2 forbids non-consensual content of real people. The FLUX.1-dev license lists it in the out-of-scope uses. The Pony Diffusion V6 XL Modified Fair AI license has its own restrictions. And every hosted CrakRevenue offer I work with bans real-person deepfakes in its terms of service. Photo-upload nudifier products are a different beast, often legally exposed under the emerging non-consensual-intimate-imagery laws, so when those searches (ai nudifier, clothes remover) land on my site I route them only to compliant fantasy-character generation, never to a photo-upload tool. That line doesn't move.

Civitai community LoRAs trained on real public figures

Same objection, same reason. LoRAs trained to reproduce a real celebrity or private person without their consent fail the consent test, no debate. I won't rank, name, or link these, even when they ship explicit output that would otherwise slot right into this list. Consent isn't optional.

SDXL base model (raw, without fine-tune)

Stability AI's SDXL base model is the foundation under pick 1 (Pony V6 XL is an SDXL fine-tune) and effectively the ancestor of the FLUX line. But raw, unmodified SDXL is trained safe-for-work and produces lousy uncensored output until you layer a fine-tune on top. The fine-tune is where the actual explicit capability lives, and Pony V6 XL is the one you download. So I point you at the fine-tune, not the base.

Mage.space, Tensor.art, Civitai Generator (cloud SD platforms)

These cloud platforms run Stable Diffusion checkpoints in your browser, no local graphics card needed. They serve a real need, but each one throttles the most explicit prompts under its content policy, and none of them carry a CrakRevenue offer I'm approved for. They're legit tools worth knowing about. They're not in my pick list for two honest reasons: there's no affiliate path, and the fit is only partial, since they restrict the exact prompts that probably brought you to this page in the first place.

How we tested

For section 1 (the open-source models), a pick had to be a Stable Diffusion checkpoint or fine-tune you can download from Civitai or HuggingFace, under a license that allows adult-content generation on consenting adult subjects, with a real creator and 100,000-plus downloads. I read every model card front to back, and the specific facts above (creator, base architecture, version date, download count, license) come straight off the official model card on Civitai or HuggingFace, verified May 27, 2026.

For section 2 (the hosted picks), a platform had to offer a publicly accessible uncensored image-gen plan, be run by a company I can verify, and carry an active CrakRevenue offer I'm approved for. Each one that cleared ran through our 8-category scoring for AI companion apps, weighted the way the scoring page spells out. The 5-prompt image-gen protocol is the loudest signal for this particular list. Current scores lean on platform marketing samples, YouTube side-by-sides under six months old, and reproducible free-trial generation where I could get it. Total spend across all this testing: $0. I use free tiers and primary policy pages, never a paid undercover account.

This is the part that matters. Most "best Stable Diffusion models NSFW" lists in this space sort by which brand pays them the most, not by which one is actually good. I don't. The hosted picks here are a deliberate mix: brands that pay well (Candy.ai and DarLink) and brands that simply fit the work better (Promptchan as the image-gen flagship, payout be damned). And when Joi pays the single highest commission in our approved catalog yet lands at #4 on this list, because the 115-character chat cap and the UK geo-block matter more to you than my margin does, that's me making the trade-off out loud instead of hiding it.

Frequently asked questions

---

Last verified May 27, 2026 · See errata log for any post-publish corrections · Editor: Alexandra Joly · Methodology · Editorial process · Affiliate disclosure

Start with the hosted pick. Try Promptchan →

Related reading

• AI image generation hub and full ranking
• Top uncensored AI image generators (sibling listicle)
• How we score AI companion apps across 8 categories
• The Promptchan deep-dive
• Candy.ai full review (8.4/10)
• The Joi scorecard
• MyLovely Ai full review (breach disclosure)
• DarLink Ai 8-category breakdown