Community Developers Unlock MOSS-Image for Adult Content Within Days of Release
A community LoRA strips the guardrails off OpenMOSS's 5B text-to-image model, turning a layout-focused DiT into an adult-content generator under Apache 2.0.
- Community LoRA Moss-image-nsfw-uncensored-lora fine-tunes MOSS-Image-SFT for adult image generation
- Released under Apache 2.0, works with both MOSS-Image Base and SFT checkpoints
- Triggered via
moss_xvstyletoken, loads through standard Diffusersload_lora_weights - Base model is a 5B DiT with FLUX.2 VAE, Qwen3.5-9B text encoder, 4K-token prompts
- Inherits MOSS-Image's text rendering and JSON layout control, but no image-to-image or reference inputs
- Signals rapid community adoption of OpenMOSS's new text-to-image platform within days of launch
A community developer has released an adult-image LoRA for MOSS-Image-SFT, OpenMOSS’s open-weight text-to-image model. The adapter page lists an Apache 2.0 license, compatibility with the Base and SFT checkpoints, and a required moss_xvstyle prompt token. Its arrival within days of the upstream release shows how quickly developers can retarget MOSS-Image for a specialized domain.
MOSS-Image provides more than general text-to-image generation. The model accepts prompts of up to 4,000 tokens, is designed to render English and Chinese text, and supports JSON layouts that place objects and text within specified bounding boxes. Those capabilities make the adapter relevant to developers who need structured compositions or rendered text alongside adult-image generation.
Three components, three jobs
| Component | Purpose | Best fit |
|---|---|---|
| MOSS-Image-Base | Pre-training checkpoint intended for further tuning | Training and adapter development |
| MOSS-Image-SFT | Supervised fine-tuned checkpoint with stronger prompt-following behavior | General inference |
| NSFW LoRA | Compact weight update that redirects either checkpoint toward adult imagery | Domain-specific generation |
A LoRA stores a relatively small set of learned weight adjustments and requires a host checkpoint at inference time. This adapter leaves the original MOSS-Image weights intact, allowing applications to load or remove it without maintaining another full model copy.
The SFT checkpoint was trained on curated data that favors safer, text-heavy compositions such as diagrams, whiteboards, and menus. The community adapter shifts that behavior toward explicit adult outputs when prompts include
This story is for Pro members
You've reached the end of the free preview. Upgrade to AlphaSignal Pro to read the full article - and everything else behind the paywall.
Comments
