There are the full official bf16 MOSS-Image SFT and Base checkpoints from Comfy dot org on Huggingface: https://huggingface.co/Comfy-Org/MOSS-Image/tree/main/diffusion_models
TA's official versions of MOSS SFT and Base are the int8 convrot variants from Comfy dot org. While fine for most purposes, if you want to use the official MOSS-Image models for the extra quality boost, use these bf16 versions instead.
MOSS-Image is a highly creative model which can produce stunning results, from artistic to photoreal.
Like Anima, it understands both tag-based (e.g., Danbooru) prompts as well as natural-language prompts.
Like Ideogram 4, it understands bounding-box prompts (via JSON), a.k.a. regional prompting, which helps avoid concept mixing and makes it easy to generate multiple-character images and complex compositions. This feature is still experimental/in beta.
Like Krea 2, its understanding of anatomy is very good.
It's excellent at rendering text.
Its composition varies highly from one seed to the next, unlike Krea 2.
It is also apparently quite easy to train.
⚠️ PLEASE NOTE: The SFT version is NOT a turbo/distilled version of the Base, needing fewer steps and a CFG of 1. MOSS-Image-SFT is the Base model fine-tuned on curated data. Its main advantage is how it follows prompts and organizes content: in the developers' comparison, SFT produces clearer text grouping and a more organized whiteboard layout. It's intended to be more useful out of the box, while Base is the foundation for further adaptation or fine-tuning (training). Both use the same architecture and inference pipeline, so SFT doesn’t inherently need fewer steps, a CFG of 1, or less compute.
To reduce the number of steps needed (and therefore use fewer credits/tokens), I recommend the following turbo LoRAs:
For SFT: 8-step data-space DMD - v2 ("Latest") - https://tensor.art/models/1057499605501007556/MOSS-Image-Turbo-LoRAs-by-Lodestones-SFT-v2-(%22Latest%22)
For Base: 8-step data-space DMD - v1 - https://tensor.art/models/1057496592581443205 or the 8-step data-space DMD2 - https://tensor.art/models/1057582086052906158/MOSS-BASE-DMD-8Step (IMO both are good, depending on the style of gen. I find DMD v1 to be slightly better for photoreal.)
*********
For optimal usage, be sure to read the official OpenMOSS user guide: https://github.com/OpenMOSS/MOSS-Image/blob/main/USERGUIDE.md
Excerpt below:
"Resolution: We recommend outputs with 1M, 1.5M, or 2M total pixels. The model can also extrapolate to arbitrary resolutions from 0.25M (512×512) to 4M (2K).
Aspect Ratio: We support arbitrary aspect ratios. Gradio’s resolution and aspect-ratio controls calculate dimensions aligned to multiples of 16. In the Python API, or when entering dimensions manually, use positive multiples of 16; invalid dimensions raise an error.
Inference Steps: Best results are achieved with ≥ 50 steps.
CFG: Default 3.0 (recommended range: 3.0–6.0). A value of 3.0 is generally sufficient for long prompts; for short prompts, we recommend increasing it as needed.
Shift: Default 5.0.
Sampler: Euler is the recommended default.
SLG (Skip Layer Guidance): We have implemented two variants:
> Standard SLG: scale 3, 0.01-1.0.
> Two-Pass Sampling SLG: We recommend uncond skip, with 22 as the recommended skip layer (0-indexed). Layers 22, 24, and 26 can all help improve image quality.
Turning off SLG will make the output image look very realistic, but it may also cause some details to collapse.
Language: We recommend English, which has the most training data. Chinese and other languages are also supported.
Negative Prompt: Based on our tests, we recommend the following effective negative tags: "worst quality, bad anatomy, watermark, jpeg artifact"
Recommended Tag Order [if using tag-based prompts]: For the best results with tag sequences, we recommend the order Quality -> Score -> Meta -> Year -> Character -> Series -> Artist -> General.
Text Layout: The model can render text and insert line breaks automatically, but the result may not always match your expectations. To improve text accuracy or control line breaks precisely, insert \n or actual line breaks directly into the prompt."