MOSS-Image

Highlights

• Native layout/box control (bboxes): set positions first, then write the description; composition control is a step above prompt-only
• Two checkpoints: SFT (instruction fine-tuned, the default in official templates) and Base (not post-trained, suited to further fine-tuning)
• Bundled prompt enhancer; a single sentence can be expanded into a full caption
• Official ComfyUI nodes already exist (search MOSS-Image in Manager) plus four official templates

Scope

• Good for: posters, storyboards, and multi-person scenes that need precise composition
• Good for: projects that want a clean Apache-2.0 commercial license
• Not for: chasing extreme photorealistic texture (its strength is layout and instruction following, not skin pores)

Tips

• Box control isn't magic: too many boxes crowd each other; start with 2–3 subject boxes
• The Base version has no post-training and produces wilder images; it's suited to further fine-tuning, not to final output
• Chinese prompts work, but official examples are mostly English; for complex layouts, English is recommended

Cautions

• The SFT and Base filenames differ by only one word; downloading the wrong one means using the wrong template

Parameters

Steps

50

CFG

3

Shift

5

Sampler

Flow-matching Euler

Resolution

Starting at 1024×1024 (official examples are 1024 square)

Text encoder

Qwen3.5-9B (set load type to moss_image)

VAE

FLUX.2 small VAE

Deployment

16GB verdict

Runs, using int8 weights plus a w4a8-quantized text encoder; that's how the official template is configured.

VRAM notes

The 5B main model isn't large; the pressure comes mainly from the 9B text encoder. First load on 16GB is slow, after that it's cached. bf16 weights also fit, but with little headroom.

Suggested starting point

Start with the official t2i template + int8 to confirm it generates images; move to the bboxes template when you need precise composition.

Comparison

Like Qwen-Image 2.1, it focuses on Chinese, but the approach differs: Qwen relies on a 7B backbone to tackle Chinese typesetting head-on, while MOSS sidesteps prompt mysticism with canvas box layout control. The cost is that it needs 50 steps and CFG 3 for good images, an order of magnitude slower than Z-Image Turbo's 8 steps. Its license is the cleanest in this batch (Apache-2.0).