Artius Moss

A photorealistic, versatile MOSS-Image merge with a carefully balanced fine-tune, LoRA and built-in turbo. Strong NSFW capabilities, sharp results in just 8–10 steps, and an optimized INT8 build that runs on GPUs with 8 GB VRAM in ComfyUI.

⚡ 8–10 steps · CFG 1 · ~7 GB · runs on 8 GB GPUs · custom text encoder & VAE included

⚖️ It's all about balance. The fine-tune and LoRAs were tested together and adjusted to work as a single model, rather than letting one dominate the result. That means no forced signature look across every image. Portraits, landscapes, fashion, fantasy, NSFW, stylized artwork — Artius Moss can handle a wide range of subjects without locking you into one aesthetic.

🎨 Built for everyday image creation, whether you're after convincing photographs, unusual artwork, or something more experimental. Describe the scene, choose your style, and see where it takes you.

✨ Why it's worth a try

• Fine-tune + LoRA merge — LoRA weights were chosen through side-by-side testing to find a balance that works in practice.

• Versatile — switch between photorealism, illustration, fantasy, and other styles without changing models.

• 🔞 Uncensored — strong NSFW capabilities, with no built-in content filtering. The model follows your prompt.

• Custom text encoder included — an uncensored Qwen3.5-9B encoder made for this model, so your prompt reaches the image without being softened along the way.

• Fast — turbo is already built in, so there's no need to load an extra LoRA.

• Fits a mid-range GPU — the optimized INT8 build takes up around 7 GB and runs on 8 GB cards.

• Ready-made workflow — drag a showcase image into ComfyUI to load the complete workflow and get started.

🔧 What's inside

• FP32 LoRA merge — all LoRAs are merged at full 32-bit precision before quantization, helping preserve fine details throughout the process.

• Calibrated layer layout — each block was evaluated for its sensitivity to quantization. The most sensitive layers retain full precision, while the others use INT8 to keep the model compact with minimal quality loss.

• Custom quantized uncensored text encoder — the custom encoder went through the same calibration: its sensitive layers stay in full precision, so prompts are understood almost exactly as by the full-size model. It can also expand short ideas into detailed prompts with ComfyUI's TextGenerate node.

⚙️ Settings that just work

• Steps: 8–10 (8 for speed, 10 for a little extra detail)

• CFG: 1 — no negative prompt needed; just describe what you want to see.

• Sampler / scheduler: euler / simple

• Resolution: 1–2 MP (e.g. 1024×1024, 912×1632)

• VAE: Flux 2 VAE (included in the files of this page)

• Text encoder: the custom encoder from this page (recommended; CLIPLoader, type moss_image). The standard MOSS-Image Qwen3.5-9B encoder also works.

📦 You'll need

• The latest ComfyUI with MOSS-Image support

• The Load Diffusion Model node

• An NVIDIA RTX 20xx or newer, with at least 8 GB VRAM

18+ only — please follow Civitai's content rules.

Built on MOSS-Image by OpenMOSS (Apache 2.0). Thanks to the original model authors, the LoRA creators, and the ComfyUI team.