CivArchive
    MOSS Image - SFT
    Preview 145881759
    Preview 145881763
    Preview 145881758
    Preview 145881761
    Preview 145881762
    Preview 145881765

    MOSS Image is a 5B parameter text-to-image DiT from the OpenMOSS team, built for accurate in-image text, layout control and long prompts. It pairs a Qwen3.5 9B text encoder with the FLUX.2 VAE, accepts prompts of up to 4K tokens in natural language, booru-style tags or a JSON layout, and renders from 512px up to 2K at any aspect ratio.

    Originally released by OpenMOSS on Hugging Face. All credit for the model goes to the OpenMOSS team. Civitai hosts a mirror so creators can generate with it on-site and train against it. Head to the original repo for the reference pipeline, the user guide and updates.

    Two versions

    • SFT is the supervised fine-tune of Base on curated data. OpenMOSS recommends it for generation: cleaner text grouping and more organised layouts.

    • Base is the undistilled base model with no post-training. OpenMOSS recommends it for fine-tuning and LoRA training. LoRAs trained on Base can also be used on SFT, with partial compatibility.

    Both run at 50 steps with CFG 3 (raise it for short prompts), flow-matching Euler, shift 5. OpenMOSS recommends English prompts; Chinese and other languages are supported. Use line breaks and quotes around text you want rendered.

    Files

    Each version carries the diffusion model in bf16 and int8, from the Comfy-Org repackage. The Qwen3.5 text encoder and FLUX.2 VAE ship in the same repo. The original Diffusers weights are at MOSS-Image-Base and MOSS-Image-SFT.

    Generating on Civitai

    Both versions are in the Civitai generator for text-to-image, with steps, CFG and a negative prompt exposed. JSON layout prompts work as plain prompt text. Image editing and reference images are not part of this release.

    License

    MOSS Image weights and inference code are released under the Apache License 2.0. Commercial use, derivatives and LoRAs are allowed; keep the licence notice with any redistribution.

    Links

    Description

    FAQ

    Comments (23)

    AeralynaiaOct 16, 2026· 2 reactions
    CivitAI

    ohh can't wait to try this one ! Thank you !!!!

    xDegenerateOct 16, 2026· 9 reactions
    CivitAI

    I'm never optimistic about newer models, but this is THE model kids, remember to use BBox, and for creators do your stuff and start creating LoRA's, this model has the potential to replace Krea 2

    Lynx2025Oct 16, 2026· 2 reactions

    yeah bbox works quite well with minimal prompt leakage

    xDegenerateOct 16, 2026· 2 reactions

    @Lynx2025 Model is insane, my Krea prompts are simplistisc AF (I don't use LLM's) but this model understands basically the most basic stuff and it gets right, it has some deformation with fingers and sometimes it gets body horror's but I believe that is a issue with the turbo LoRA, but with time it will be a great model.

    GodGooner69Oct 17, 2026· 5 reactions
    CivitAI

    Yeah this is gonna replace Krea 2 probably. Hopefully Krea 3 takes notes otherwise it'll be irrelevant with this model's release.

    elevendrOct 17, 2026· 3 reactions
    CivitAI

    What is Moss Image and how dose it comepare to Krea 3/2 and Qwen Image 2.1?

    prestonvindeuxOct 17, 2026· 3 reactions
    CivitAI

    Interesting. Quite costly at like 40 buzz / image. But this is without Turbo LoRAs.

    reakaakaskyOct 17, 2026· 10 reactions
    CivitAI

    RIP Flux3 klein for sure .... and probably Krea2 and Anima (new anime dataset cutoff ~05.2026).

    prestonvindeuxOct 17, 2026· 2 reactions

    Krea2 still good for the near future. 12 buzz/image with Turbo. A lot of LoRAs and checkpoints. It'll take a few months for MOSS to truly unseat it.

    aiklaaniOct 17, 2026
    CivitAI

    Cant make it work localy on Comfyui, it just not see config.json in transformer/

    vamorandOct 17, 2026

    comfyui works fine. update comfy to latest build instead of stable, or just wait a bit for a new stable update

    TalancaOct 17, 2026· 12 reactions
    CivitAI

    WOW, fully uncensored from the factory

    DepurtureOct 17, 2026· 4 reactions
    CivitAI

    One day I'll see an open weight diffusion transformer this small that looks good and edits too.

    vamorandOct 17, 2026· 2 reactions
    CivitAI

    It knows many characters and concepts (best NSFW base model yet), but prompting is hard: results are unstable with many artifacts. Maybe a good skill with decent LLM can make it right, utilizing it's 4K context window (Krea 2 only has 512)

    dawn66666666Oct 17, 2026· 7 reactions
    CivitAI

    This is, to date, the most knowledge-rich open-source text-to-image model (for realistic).

    It is capable of handling both anime and realistic styles, and it is uncensored. Although there are occasional issues with limb structure, I believe it will only get better in the future.

    JFTCOct 17, 2026· 4 reactions
    CivitAI

    This is a nice little model, size wise, cause in terms of realism it can give Qwen2.1 a run for their money. I still like Anima more but I've heard they're working on Anime targeted version.

    natasha_26Oct 17, 2026
    CivitAI

    A new king is here🙌.Moss blow out of the water Krea 3 Klein and I’d say it’s slightly superior to the Krea 2.

    JFTCOct 17, 2026· 1 reaction

    I don't think it's superior to Krea2 honestly, but definitely a nice mix of Flux and Anima. If they release MOSS XL or something like a 12B parameter model, then yeah possibly. The license is very good as well.

    BonerSoupOct 17, 2026
    CivitAI

    This model is pretty damn incredible.

    it can handle complex poses pretty well, and is pretty creative with compositions and styles.

    mphobbitOct 17, 2026· 4 reactions
    CivitAI

    MOSS Image extension for Neo by eduardoabreu81:
    https://github.com/eduardoabreu81/moss-image-forge-neo

    JFTCOct 17, 2026

    That was quick.

    r3cOct 17, 2026· 2 reactions
    CivitAI

    This model has potential.

    Out of the box, its uncensored and shows decent prompt adherence to both tags and natural language. It's got bounding boxes which enable more complex compositions as well.

    I've not got it to work with artist style tags, so I'm guessing they were pruned from the dataset.

    With a good finetune, this would overshadow Anima, however, I'm interested to know what differences in training methodology would this model require. Guess it's time to take a look at the paper.

    infernahermit846Oct 17, 2026
    CivitAI

    Generations are too Slow (+100seconds). needs turbo lora.
    Quality is really good tho, one of if not the best.

    Checkpoint
    MOSS Image
    by CivitaiOfficial
    Download (Beta) View on CivitAI

    Details

    Downloads
    1,551
    Platform
    CivitAI
    Platform Status
    Available
    Created
    10/17/2026
    Updated
    10/17/2026
    Deleted
    -

    Files

    mossImage_sft_txt_int8.safetensors

    Size:
    9.08 GB
    SHA256:
    32e3787472f708fa5f83624acbf5ef23f7fc89c439e05065ea57cfe5a97f025a

    Mirrors

    HuggingFace (1 mirrors)
    CivitAI (2 mirrors)
    ModelScope (1 mirrors)

    mossImage_sft_bf16.safetensors

    Size:
    9.47 GB
    SHA256:

    Mirrors

    mossImage_sft_pruned_int8.safetensors

    Size:
    5.04 GB
    SHA256:

    Mirrors

    mossImage_sft_txt_int4.safetensors

    Size:
    6.84 GB
    SHA256:
    4cc6a1543ad2061806da46f35f90cadb4801dcba246c44782cf42ec0feb6407c

    Mirrors

    HuggingFace (1 mirrors)
    CivitAI (1 mirrors)
    ModelScope (1 mirrors)

    flux2-vae.safetensors

    Size:
    320.64 MB
    SHA256:
    868fe7b343cc8f3a19dbcfcafbc3d5f888802be3f89bd81b65b3621a066ce8f3

    Mirrors

    HuggingFace (136 mirrors)
    CivitAI (10 mirrors)
    ModelScope (8 mirrors)
    ModelScope CN (4 mirrors)