About MOSS

MOSS-Image is a promising model for furry art: SFT draws well-known species cleanly and follows long, detailed prompts closely. The landscape around it is still very young, with few fine-tunes and LoRAs so far, and artist tags have no visible effect. For now the style comes from describing the medium, as in the examples below, on top of SFT's own polished look.

MOSS-Image understands plain English, booru-style tags and a mix of both. Tags go in a fixed order: quality, score, meta, year, character, series, artist, general. English works best, Chinese and other languages also work.
Long, detailed prompts work well, up to 4K tokens. For a long prompt CFG 3 is enough, a short prompt may need more, up to 6.

MOSS-Image is a 5B parameter model by the OpenMOSS team, released as open weights in October 2026. It reads the prompt with the Qwen3.5 9B text encoder and uses the Flux.2 VAE. It can also write text into the image.

The two versions:
SFT: Base fine-tuned on curated data, for making images.
Base: the raw model without post-training, meant for training LoRAs and fine-tunes.
Both run at 50 steps and CFG 3. A faster, distilled version is announced but not out yet.

The checkpoints are about 9.5 GB in bf16 and 5 GB as int8. The text encoder qwen3.5_9b is about 18 GB in bf16, 9 GB as int8 and 7 GB in 4 bit; the VAE is flux2-vae. All of them are in the Comfy-Org/MOSS-Image repository on Hugging Face.

The sampler grids below use six samplers and five schedulers. Karras is in them on purpose: it blurs the image with every sampler, and its row shows that. The others were tested on SFT:

• UniPC, Res Multistep, DEIS: at 50 steps all three render nearly the same image, so only UniPC is in the grid. Euler a and the two SDE samplers add new noise at every step and draw a different picture than the others.
• LCM: left out. It is made for distilled few-step models and comes out flat and washed out.
• KL Optimal: left out. It drifts towards the same soft, fluffy cat as Karras, only less blurry.
• Linear Quadratic, Bong Tangent: left out. Both work at 50 steps, but give practically the same image as Simple.