AI models tested for NSFW

Alibaba’s Qwen3.8 generation (August 2026): the 27B checkpoint is Apache 2.0 with vision and a 262K context, and abliterated “uncensored” builds appeared within a day. The likely next base for roleplay fine-tunes.

ByteDance’s flagship image model (August 2026): text-to-image, image editing, up to 10 reference images, strong text rendering. Available only through ByteDance and partner APIs with content moderation.

Lightricks’ 22B open-weights video model (August 2026): 10-second 720p image-to-video in seconds on datacentre GPUs, 4K HDR, native multishot, runs quantised on a 24–32 GB card. No built-in filter; NSFW LoRAs from LTX-2.3 carry over.

Meta’s return to open weights (August 2026): a 30B Apache 2.0 model built for local agents, vision input, 131K context, runs on one consumer GPU at 4-bit. Not tuned for roleplay; the community will decide whether it becomes a base.

ByteDance’s newest Seedance tier: clips up to 30 seconds at up to 1080p with multishot and audio. Filtered on official surfaces; runs with adult content enabled on MyBabes.

Alibaba’s next Wan generation, offered through the DashScope API with 2–30 second clips up to 1080p. Hosted only so far; open weights not yet released.

ByteDance’s flagship video model. Heavily filtered on official apps, available with adult content enabled on MyBabes.

OpenAI’s image model (April 2026), 2K output, reasoning mode, strong text rendering. Refuses explicit nudity and sexual content at the model level; the API moderation setting shifts thresholds but does not remove them.

Google’s open-weights Gemma 4 family and the uncensored roleplay fine-tunes built on it (26B class), which power MyBabes’ persona chat.

Alibaba’s hosted Wan generation (API-only since April 2026; open weights stop at 2.2). Strong body physics and prompt adherence over 5–15 s clips. Runs with adult content enabled on MyBabes through a provider route.

Alibaba’s open-weights Qwen 3.5 and the creative-writing fine-tunes on it (27B Writer), used on MyBabes as the premium roleplay model.

Popular hosted video model with strong motion quality. Official app blocks nudity and explicit content.

Real-time image generation and enhancement. Vendor filters explicit prompts; MyBabes runs an uncensored pipeline.

xAI’s image model, available through the x.ai API. Permissive on suggestive content compared with other closed vendors, but explicit output is filtered.

Open-weights image model family. Uncensored community fine-tunes and LoRAs exist; hosted versions are filtered.

The previous-generation open-weights Wan model, self-hosted on MyBabes as the “MyBabes AI NSFW” video option with LoRA mods for specific acts and camera styles.

Open-weights text-to-speech with voice cloning from a short reference, self-hosted on MyBabes for character voice messages.

The SDXL-derived family that succeeded Pony for anime and stylised adult art, with stronger character knowledge and cleaner tag adherence; realism merges exist too.

The SDXL-derived family trained on tagged art datasets that became the standard base for stylised and semi-realistic adult image checkpoints.

The workhorse of self-hosted NSFW image generation. Needs a GPU, a UI and the right checkpoints.

The leading hosted TTS service, used on MyBabes for selected voices and sound effects; its terms restrict explicit content.