If you want to generate uncensored anime locally in 2026 — meaning a .safetensors file on your own GPU, not a cloud SaaS that filters your prompt — there are roughly ten checkpoints that actually matter. The rest is noise: orphaned 2023 SD1.5 models, reskinned mergers with no community, and closed-API products that quietly add a moderation layer between you and the diffusion step.
This list is the short answer. We go from the highest-priority pick to the most niche, sorted by what your GPU can actually run. Every model below is downloadable, weights-open, and has a real Civitai or HuggingFace footprint. We tested with ComfyUI on an RTX 4070 (12GB) and an RTX 3090 (24GB), so the VRAM tiers are honest, not theoretical.
We assume you already know what a checkpoint is and roughly how SDXL or FLUX models load. If not, the four-minute setup section below catches you up.
What "Uncensored Anime" Actually Means In 2026
The phrase gets used loosely, so let's nail it down.
"Uncensored" in the open-weights world means three things stack together. First, the base model was trained on Danbooru-tagged or Gelbooru-tagged data — that means anatomy, character names, costume details, and explicit content are all known concepts the model can compose. Second, there is no per-frame moderation classifier intercepting your output the way Bing Image Creator or Midjourney do. Third, the license permits NSFW generation rather than carving it out as a forbidden use.
This is the opposite of how SaaS anime generators work. NovelAI's image diffusion API, getimg.ai, Tensor.art's hosted generation, Civitai's on-site generator — all of those run filters on top of weights you cannot inspect. Your prompt goes in, classification runs, output gets blurred or rejected if it tripped a threshold. The model is uncensored, the product is not.
The picks below are checkpoints. You download a 6.5GB to 13GB file, drop it into ComfyUI or Forge or reForge, and the only filter is the one you write into your negative prompt. Danbooru-tag-aware means typing 1girl, smile, school uniform, classroom actually does what you'd expect. Score-tag-aware (Pony family) means typing score_9, score_8_up, score_7_up, source_anime is the standard quality boost.
Audience here is the local-generation user. If you want a hosted product, this list is not for you.
The Pony / Illustrious Split
The single most important fact about anime image generation in 2026 is that there are two competing prompt-syntax universes, and they are not interchangeable.
Pony family descends from Pony Diffusion V6 XL, released by AstraliteHeart in late 2023 on top of SDXL. It rebuilt SDXL's text encoder around a score_X quality tagging system and source_anime / source_furry / source_pony / source_cartoon content sources. To get good output, you start every positive prompt with something like score_9, score_8_up, score_7_up, score_6_up, source_anime and then describe the scene in tag-style. Booru tags work, but the score prefix is non-negotiable for quality. Every Pony fine-tune and merger in this list inherits this syntax — Pony Realism, Babes by Stable Yogi, the Pony side of any Pony/Illustrious hybrid.
Illustrious family descends from Illustrious XL v0.1, released by OnomaAI Research in December 2024. It was trained on roughly 7.5 million Danbooru-tagged images on an SDXL base, with no score-tag system. The prompt syntax is plain Danbooru: 1girl, blue hair, twintails, school uniform, classroom, sunlight — exactly the format you copy out of a Danbooru post. NoobAI XL, Hassaku, WAI, Prefect, Cat Tower — all Illustrious-derived, all use plain Danbooru tags.
Mixing the syntaxes does not work. Score-tag prompts on an Illustrious model produce noise. Plain Danbooru tags on a Pony model produce mediocre composition because you skipped the quality conditioning. The first decision you make before downloading anything is which universe you live in.
The April 2026 split. Last month dropped two major releases. AstraliteHeart shipped Pony V7 on the AuraFlow base — much higher fidelity, but it broke LoRA compatibility with the entire V6 ecosystem. Years of community LoRAs, every character mod, every style mod — none of it loads on V7. Adoption has been slow because the cost-of-switching is brutal. OnomaAI shipped Illustrious 2.0 the same month and made the opposite call: stay on SDXL, preserve LoRA continuity. Existing Illustrious LoRAs work. Existing mergers work. The community kept moving without a fork.
Most users in May 2026 are still on the V6 / v0.1 generation. The mergers below reflect that. We mention V7 and 2.0 where relevant, but the picks are the checkpoints people are actually downloading right now.
How To Run These (4-Minute Setup)
Install ComfyUI (the standalone Windows portable build, or git clone + python main.py on Linux/Mac). Drop your .safetensors file into ComfyUI/models/checkpoints/. Start ComfyUI, load the default workflow, change the checkpoint loader node to point at your downloaded model. Type a prompt. Hit Queue.
That's it. No login, no cloud round-trip, nothing leaves your machine.
For Pony-family models, your default positive prefix is score_9, score_8_up, score_7_up, source_anime, followed by your actual prompt. For Illustrious-family models, just write Danbooru tags. Negative prompts: worst quality, low quality, jpeg artifacts, blurry, watermark is a fine default and applies to both families.
Sampler: Euler A or DPM++ 2M Karras at 25–30 steps for SDXL-based models. CFG 5–7. Resolution 1024x1024 native, or 832x1216 for portraits, 1216x832 for landscape. Anything outside the SDXL aspect-ratio buckets degrades quality.
If you want a UI that's less node-graph and more A1111-style, Forge or reForge load the exact same checkpoints with a traditional sliders interface. Same .safetensors files, different frontend.
The Picks (Ranked By VRAM)
Tier 1: 8GB VRAM (RTX 3060 / 4060 / older 2070 Super)
Every SDXL-based pick below runs on 8GB VRAM at 1024x1024 with reasonable speed. This is the entry tier and where most of the audience sits.
1. Pony Diffusion V6 XL — the load-bearing base
Pony V6 is the model that defined uncensored anime SDXL for two full years. AstraliteHeart trained it on a curated Danbooru + e621 + custom dataset with an aggressive score-tagging schema, and the result was the first anime SDXL fine-tune that handled NSFW anatomy reliably without LoRAs. Every Pony-family pick on this list is downstream of V6.
What it's good at: dynamic poses, NSFW anatomy, character consistency across a session, and a slightly stylized "Pony aesthetic" that some users love and some users layer style LoRAs to escape. Composition is strong even on under-described prompts. Multi-character scenes hold up surprisingly well for an SDXL model.
Prompt syntax: full score-tag prefix is mandatory. score_9, score_8_up, score_7_up, score_6_up, source_anime, 1girl, .... Skipping it is the single most common reason new users get bad output. License is permissive for personal use; check the Civitai license page before any commercial workflow because the terms have specific clauses around model resale and training.
2. Illustrious XL v0.1 — the new dominant base
If Pony V6 is the past, Illustrious v0.1 is the present. OnomaAI's December 2024 release trained on 7.5M Danbooru-tagged images and immediately became the merger base most 2025 anime checkpoints built on. The plain Danbooru-tag syntax is much friendlier for new users — no quality prefix, just describe what you want.
What it's good at: tag fidelity, character recognition (Danbooru artist tags often work), anime-native aesthetics that don't read as "Pony output." NSFW anatomy is competent but less aggressive than Pony out of the box; the mergers below address that. Composition is slightly weaker than Pony for complex multi-subject scenes but the rendering quality per subject is higher.
Prompt syntax: plain Danbooru tags, masterpiece, best quality as a soft quality prefix is optional but commonly used. License is research-permissive with commercial-use clauses; OnomaAI explicitly allows derivative mergers, which is why the Illustrious ecosystem exploded so fast.
3. Hassaku XL Illustrious — the clean-anime merger
Hassaku has been a long-running anime merger series since the SD1.5 era. The Illustrious version is the one to download in 2026. Creator: Cyberdelia. The merger leans toward clean linework, saturated color, and a slight cel-shading bias — closer to a modern TV anime look than the painted-illustration look of Animagine.
What it's good at: SFW and lightly NSFW anime portraits, character art that reads as "official artwork" rather than fan art, color palettes that don't drift muddy. NSFW capability is present but the aesthetic stays clean — less suited for explicit content than Pony Realism or Babes, better suited for borderline / suggestive work.
Prompt syntax: plain Danbooru, no score tags. License inherits from Illustrious permissive terms.
4. WAI Illustrious SDXL — the all-rounder merger
WAI is the merger that gets recommended on /h/ threads as "just use this if you don't know what to pick." It's Illustrious-based, balanced between SFW and NSFW, balanced between styles. Less opinionated than Hassaku, less specialized than Cat Tower. The trade-off is that it does nothing badly and nothing exceptionally — a pure mid-curve recommendation.
What it's good at: prompt adherence across a wide range of subjects, decent NSFW anatomy without prompt gymnastics, character LoRA compatibility (almost any Illustrious LoRA loads cleanly on top). Good first download for users who don't know yet what aesthetic they want.
Prompt syntax: Danbooru tags. License: Illustrious-permissive.
5. Prefect Illustrious XL — the painterly one
Prefect is the Illustrious merger that pushes toward a softer, more painterly aesthetic — closer to digital illustration than to TV anime cel-shading. Creator targets a "high-quality artbook" feel and largely succeeds. If Hassaku is a clean-line anime DVD, Prefect is a print artbook.
What it's good at: portraits, fantasy scenes, characters with elaborate costumes where the rendering complexity benefits from the softer style. Less appropriate for action scenes or comedic anime where the painterly bias slows the read.
Prompt syntax: Danbooru tags, often benefits from explicit style tags like oil painting, pastel, soft lighting. License: Illustrious-permissive.
6. Nova Anime XL — the hybrid
Nova Anime is the most popular Pony/Illustrious hybrid checkpoint — a merger that genuinely tries to inherit useful properties from both families. The result is a model that responds to plain Danbooru tags but also benefits from a partial score prefix. It's the "fence-sitting" pick and that's intentional, not a flaw.
What it's good at: users who switch between Pony-style and Illustrious-style prompting and don't want to maintain two checkpoints. NSFW anatomy is solid, slightly closer to Pony's aggression than to vanilla Illustrious. Aesthetics lean modern-anime.
Prompt syntax: works with either family but optimal results come from masterpiece, best quality, score_9, source_anime, then Danbooru tags. License: hybrid-derivative, check the Civitai page before commercial use.
7. Animagine XL — the pre-Illustrious base
Animagine XL is the SDXL anime fine-tune that dominated before Illustrious shipped. Cagliostro Lab's training run was the first widely-adopted Danbooru-tag anime SDXL model and a lot of users still prefer its rendering to Illustrious for specific aesthetics — particularly more traditional / 2010s-anime looks.
What it's good at: classic anime aesthetics, character recognition for older anime characters whose Danbooru tag presence pre-dates Illustrious's training cutoff, scenes where Illustrious's modern bias feels too "AI art." Mostly SFW-tuned but handles light NSFW.
Prompt syntax: Danbooru tags with masterpiece, best quality, very aesthetic as the standard quality prefix. License: research-permissive with commercial-use clauses, well-documented on the Cagliostro repo.
8. Cat Tower NoobAI XL — the Danbooru-purist pick
NoobAI XL is a separate SDXL anime base trained on the Danbooru noobai dataset, distinct lineage from Illustrious. Cat Tower is the most-downloaded NoobAI merger checkpoint, and it's the pick for users who want maximum tag fidelity at the cost of some aesthetic refinement.
What it's good at: niche character recognition (NoobAI's training set captured tags that Illustrious missed), tag-precise composition, NSFW anatomy with minimal prompt engineering. The aesthetic is rawer and less polished than Hassaku or Prefect, which some users prefer as a feature.
Prompt syntax: pure Danbooru tags, often with explicit masterpiece, best quality, newest prefix. License: NoobAI-permissive, broadly similar to Illustrious terms.
Tier 2: 12GB VRAM (RTX 3060 12GB / 4070 / 4070 Super)
The 12GB tier opens up the same SDXL picks at higher resolutions (up to ~1536x1536 with VAE tiling) and gives you headroom for upscaling workflows. Two picks below specifically benefit from this tier.
9. Pony Realism — anime composition, photoreal output
Pony Realism is the Pony V6 fine-tune that crosses the streams: Pony's compositional instincts and tag fidelity, but the rendering trained toward photorealism. The result is a model that takes anime-style prompts (score_9, source_anime, 1girl, twintails, ...) and returns photoreal output — like a photorealistic cosplayer of an anime character.
What it's good at: photoreal output where you want anime-style composition control, NSFW with the anatomical precision Pony is known for but a non-anime aesthetic, "realistic anime cosplay" workflows. Not the right pick if you actually want anime aesthetics — for that, use Pony Diffusion V6 XL directly.
Prompt syntax: Pony score-tag system. License: Pony-derivative, inherits V6 terms. Runs comfortably on 8GB but benefits from 12GB headroom for larger output and upscale passes.
10. Babes by Stable Yogi (Pony) — the NSFW Pony specialist
Babes is the most-downloaded NSFW-specific Pony V6 fine-tune. Stable Yogi's training run pushed harder on anatomy, body diversity, and explicit content than vanilla V6, and the result is the de-facto Pony NSFW pick when V6 itself isn't aggressive enough.
What it's good at: explicit NSFW anime, body type diversity, anatomy precision under heavy occlusion or unusual poses. Less suited for SFW work — the model has a strong NSFW prior that bleeds into innocuous prompts unless you negative-prompt against it.
Prompt syntax: full Pony score-tag prefix, mandatory. License: Pony-derivative. Runs on 8GB but benefits from 12GB for batch workflows.
Tier 3: 16-24GB+ VRAM (RTX 4080 / 4090 / 3090 / 5080+)
Honest answer: if you have a 4090, the SDXL anime picks above are still the best place to start, because the ecosystem is overwhelmingly SDXL. The high-VRAM tier opens up FLUX-anime-merge experiments and Pony V7 (AuraFlow base) at usable speed, but the community library on those is still nascent in May 2026. Wait one to two more quarters before committing your workflow to FLUX-anime unless you specifically want to be early.
For now: high-VRAM users should run the SDXL picks at 1536x1536 native, then upscale-pass to 3072x3072 with tiled diffusion. That workflow produces output quality the FLUX-anime mergers cannot yet match.
Side-by-side Table
| Model | Base | NSFW | VRAM | Prompt syntax | Best for |
|---|---|---|---|---|---|
| Pony Diffusion V6 XL | SDXL/Pony | Yes | 8GB | score_X tags | Reliable NSFW anime base |
| Illustrious XL v0.1 | SDXL | Light | 8GB | Danbooru tags | The new default base |
| Hassaku XL Illustrious | Illustrious | Light | 8GB | Danbooru tags | Clean cel-shaded anime |
| WAI Illustrious SDXL | Illustrious | Yes | 8GB |
Choose by use case
SFW anime (portraits, illustration, character art): Hassaku XL Illustrious or Prefect Illustrious XL. Both are Illustrious-based, both lean clean and high-quality, both work with the broadest LoRA library. Hassaku for cel-shaded TV-anime aesthetics, Prefect for painterly artbook aesthetics.
NSFW anime (explicit, anatomy-precise): Babes by Stable Yogi (Pony) for the most aggressive NSFW prior, or Pony Diffusion V6 XL for general NSFW with broader aesthetic range. Cat Tower NoobAI is the Illustrious-side pick if you prefer plain Danbooru syntax and want to avoid score tags entirely.
Photoreal anime (cosplay-style, realistic output, anime composition): Pony Realism. There is no real second choice in this niche — most other "realistic anime" mergers are reskins of the same base.
Furry / anime-furry hybrid: Pony Diffusion V6 XL natively, using source_furry or mixing source_anime, source_furry in the prompt. The Pony base was trained on substantial e621 data alongside Danbooru, which is why it dominates this niche. Illustrious-family models are pure-anime and do not handle furry well without dedicated LoRAs.
Niche character recognition (older anime, obscure characters): Cat Tower NoobAI XL for Danbooru-tag depth, or Animagine XL for pre-Illustrious-cutoff characters.
What We Excluded
Closed APIs. NovelAI's image diffusion endpoint, getimg.ai's hosted generation, Tensor.art's on-site generator, Civitai's hosted generator, Leonardo, Midjourney, Bing Image Creator, Adobe Firefly, OpenAI image gen. All run moderation between your prompt and your output. None are downloadable. None belong on a local-generation list, regardless of how good their underlying weights might be.
Filter-by-default models. SDXL base, SD1.5 base, SD3, FLUX.1 dev (without anime fine-tunes), Stable Cascade. These have no anime prior worth mentioning and their training data was deliberately scrubbed of explicit content. They're general-purpose foundation models — not anime generators.
SD1.5 anime checkpoints. Counterfeit V3, AnythingV5, MeinaMix, Hassaku V1.x. All were excellent in 2023–2024 but the SDXL ecosystem has surpassed them on every axis: resolution, anatomy, prompt fidelity, LoRA library, character recognition. Keep them if you have nostalgia. Don't start there in 2026.
Pony V7 and Illustrious 2.0 (April 2026). Both shipped last month. Illustrious 2.0 stayed on SDXL and preserved LoRA continuity, so adoption is happening organically — most of the mergers above will release 2.0-based variants over the next quarter and we'll re-rank then. Pony V7 jumped to AuraFlow and broke LoRA compatibility, so the ecosystem is currently fragmented and most users are waiting for V7-native mergers and LoRAs to mature before switching. Neither is excluded permanently — they're just not the right recommendation for May 2026.
