NSFW AI Model Leaderboard

The next generation of FLUX raises the ceiling for open-weights image generation. Prompt adherence and coherence jump from great to borderline unfair.

Open-weights TTS with 20+ emotion controls. The ability to dial in breathy, whispered, or intense delivery makes it a natural fit for companion voice.

The benchmark everyone chases. Prompt adherence and anatomy are absurdly good for an open model.

Four steps and you're done. Quality takes a hit, but nothing else touches this speed.

Cinematic lighting out of the box. Handles full-body compositions without the usual SDXL hand disasters.

xAI's frontier model with the loosest content policy among the majors. Will write explicit scenes without much pushing, though it still has some topic limits.

Sao10K's 8B entry built explicitly for NSFW. Runs anywhere, refuses nothing, and writes better explicit content than models three times its size.

Tiny model that outperforms its weight class. 82M parameters running on CPU with quality that rivals cloud TTS. Open weights, no content filter, absurdly cheap.

The NSFW anime checkpoint that rewired Civitai. Tag-based prompting gives you surgical control over poses and anatomy.

Photorealism that fools people. Skin texture and lighting are a cut above anything else on SDXL.

Community derestriction of GLM-4.6 that tops the UGI rankings. Massive MoE architecture delivers frontier quality with zero content filters.

The OG uncensored roleplay model that proved small models can write. Runs on a potato and still delivers surprisingly coherent NSFW scenes.

The workhorse that launched a thousand fine-tunes. Still the widest LoRA ecosystem by far.

Built for anime and illustration with a massive tag vocabulary. The Danbooru training data shows in the best way.

Trained explicitly for NSFW with zero safety filters baked in. Anatomy accuracy for explicit poses is unmatched in the SDXL family.

Budget king from Alibaba. Quality punches near FLUX.2 at a fraction of the cost, and open weights mean you can strip the safety filters yourself.

The open-source video model that actually competes with closed APIs. Motion coherence over 5+ seconds is wild.

Major upgrade over LTX-Video with better temporal consistency. Open weights and Apache license make it the fastest path to uncensored video generation.

Punches above its weight class for 8B parameters. Runs on consumer GPUs and still writes convincing scenes.

DeepSeek's latest MoE giant with open weights. General intelligence rivals GPT-4 class models, and the open nature means no hard content blocks.

The community NSFW roleplay standard. Sao10K's fine-tune delivers rich, flowing prose and handles complex multi-character scenes without losing the thread.

Built on Qwen2 with training data curated for creative fiction. Writes with a rhythm that feels authored, not generated. Handles kink negotiation in character.

Mistral's 12B sweet spot balances quality and speed. The base model has mild guardrails that community fine-tunes strip out easily.

Low-latency voice cloning that handles breathy and expressive tones well. The open-source voice model to watch.

Emotion and style controls baked into the architecture. You can shape tone without prompt hacking, and open weights mean no guardrails on what you generate.

Generates dialogue between multiple speakers from a script. Natural turn-taking and emotion tags make it interesting for AI companion voice scenarios.

The 70B merge that keeps the uncensored roleplay community fed. Prose quality and character consistency are hard to beat at this size.

Built for NSFW chat with zero guardrails. Keeps character voice consistent across long sessions without drifting.

Purpose-built Mistral merge for unrestricted creative writing. Prose quality at 123B is a real step up from the 70B class, and it refuses nothing.

The UGI #4 open-weights model. Named like malware, writes like a bestselling author. Zero refusals and strong long-context performance.

The 671B MoE model that shocked the industry with GPT-4 class performance at open-weights pricing. Mild content filters in the official API, but self-hosted removes them.

Community merge on Llama 3.3 that scores high on both general intelligence and willingness. Prose quality is polished for a merge.

The voice quality benchmark everything else gets measured against. 70+ languages, incredible expressiveness, but content policy blocks explicit audio.

~75ms latency makes this the speed king. Quality trades off slightly for real-time responsiveness, perfect for live AI companion voice.

Diffusion-based TTS that produces remarkably natural prosody. Academic origin means no commercial agenda and no content filtering.

Context-sensitive speech model that maintains voice character across long conversations. Designed for sustained dialogue, not just isolated utterances.

Clean linework and consistent character faces. Better at SFW anime than explicit content, but it handles both.

Merge focused on stylistic prose with a slant toward dramatic, literary NSFW writing. Less casual than MythoMax, more novel than chatbot.

Voice cloning TTS that supports 17 languages. Coqui shut down, but the model lives on. Still a solid choice for self-hosted voice cloning with no restrictions.

Strong prompt following and text rendering, but the NSFW fine-tune scene hasn't caught up yet.

Commercial FLUX.2 variant with safety filters baked in. Extraordinary quality but locked down for adult content.

Cinematic motion quality with good temporal consistency. NSFW fine-tunes are still early but promising.

Lightweight and fast for a video model. Quality trails the heavyweights, but iteration speed makes up for it.

Incremental upgrade that smooths out motion artifacts from the original. NSFW fine-tunes are gaining traction in the community.

Incremental open-weights update to Wan 2.1. Community NSFW fine-tunes carry over from the original, and quality got a small bump.

Meta's workhorse that spawns most of the best NSFW fine-tunes. The base model resists explicit content, but the architecture is the foundation of the uncensored LLM ecosystem.

Alibaba's challenger that matches Llama 3.1 70B on benchmarks while being more permissive out of the box. Increasingly popular as a base for uncensored merges.

29-language TTS with natural prosody. The previous gen workhorse before v3 dropped, still widely used for its stability.

MiniMax's HD tier rivals ElevenLabs on raw quality. The latency is higher, but the output sounds startlingly human.

Speed-optimized MiniMax variant. Quality dips slightly from HD but inference speed makes it viable for real-time applications.

Describe the voice you want in natural language and it generates matching speech. Novel approach that opens creative possibilities for character voices.

Zero-shot voice cloning from short audio samples. English-focused but handles emotional range reasonably well for its size.

OpenAI's native image generation through GPT-4o. Text rendering and compositional understanding are a tier above everything else, but strict content filters kill the NSFW use case.

Still the aesthetic benchmark that everything gets compared to. Art direction and composition are second to none, but Discord-only access and zero NSFW tolerance limit its reach.

Mistral's flagship API model with strong multilingual performance. Content filtering is present but less aggressive than OpenAI or Anthropic.

Six built-in voices, low latency, dead simple API. Not the most expressive option but reliable and fast, with strict content filtering.

Google's latest image model with photorealistic output that rivals GPT Image. Completely walled off from adult content through Vertex AI safety layers.

Solid text-to-video with decent motion, though it struggles with complex multi-subject scenes.

xAI's video generation leads on image-to-video benchmarks. Slightly less filtered than competitors but still blocks explicit content.

Google's Studio voices sound polished and professional. Large language support, but the content policy is as strict as you'd expect from Google.

Gemini's multimodal architecture applied to speech. Fast and contextually aware, but filtered through Google's safety layers.

Generates speech, music, and sound effects from text prompts. Quality trails purpose-built TTS models, but the creative flexibility is unmatched and it's completely open.

Design-focused model that nails text rendering inside images. Built for branding and illustration, not your goon cave.

Strongest text-in-image generation on the market. Logos, posters, and typography render clean, but the safety rails block adult prompts hard.

Consumer-friendly image generation with a polished web interface. Quality is competitive but the walled-garden model means no path to NSFW.

The current king of AI video. Motion quality and temporal consistency at 1080p are absurd, but zero chance of NSFW output through their API.

Newcomer that's eating Kling's lunch on quality benchmarks. Priced lower too, but still fully filtered for safety.

Previous gen Kling that's still extremely competitive. Turbo mode runs faster than most alternatives at similar quality.

Cohere's RAG-optimized model with open weights. Strong for factual writing and search grounding, but not built for creative NSFW and it shows.

Higher fidelity version of TTS-1. Slower but noticeably cleaner audio, especially on sibilants and breathy passages.

The previous gen OpenAI image model. Prompt following is strong for its era, but it's been lapped by GPT Image and FLUX.2.

Runway's latest pushes cinematic quality and camera control. Hollywood uses it; your goon cave cannot.

Google's video model with native audio generation. Cinematic motion and scene transitions are top tier, locked behind Vertex AI safety.

The proprietary API version of Wan pushes quality further than the open release. Filtered through Alibaba's cloud safety, so NSFW is blocked.

OpenAI's video flagship. Impressive scene understanding and physics simulation, but the most expensive option by far and aggressively filtered.

Consistently strong across both text-to-video and image-to-video. Good value for the quality, but the content policy is strict.

Chinese-made video model that ranks surprisingly high on global benchmarks. Image-to-video quality is a standout.

Luma's latest handles 3D-consistent motion well. Camera movements feel natural, but the content filter catches anything remotely suggestive.

Consumer-friendly video generation with fun effects and lip-sync. More playful than cinematic, and locked down tight on content.