MOSS-Image: Open 5B Image Model with Layout Control
ComfyUI Wiki
MOSS-Image is OpenMOSS' open 5B text-to-image family with layout JSON control, accurate English and Chinese text rendering, and ComfyUI support in core.
M
MOSS-Image
Text-to-ImageLayout ControlText RenderingOpen WeightsOpen 5B text-to-image family from OpenMOSS. A single-stream diffusion transformer in the FLUX.2 latent space with a Qwen3.5-9B text encoder, built for legible text rendering and JSON layout control: describe the scene, then place text boxes and objects by bounding box. Ships as a base checkpoint for fine-tuning and a supervised fine-tune for everyday generation.
| Developer | OpenMOSS |
| Release Date | 2026-10-16 |
| Architecture | 5B single-stream DiT (42 layers) in the FLUX.2 latent space, Qwen3.5-9B text encoder, FLUX.2 VAE |
| Parameters | 5B |
| License | Apache-2.0 |
| Resolution | 1M, 1.5M or 2M pixels recommended, 0.25M to 4M supported, multiples of 16 |
| Sampling | 50 steps or more, CFG 3, shift 5, negative-branch skipping at layer 22 |
| ComfyUI | Core support merged in Comfy-Org/ComfyUI PR #17095, repackaged weights in Comfy-Org/MOSS-Image |
Models
| Model | Description | License | Released |
|---|---|---|---|
| MOSS-Image Base | Pre-training checkpoint with the full training signal, intended for LoRA and fine-tuning work | Apache-2.0 | 2026-10-16 |
| MOSS-Image SFT | Supervised fine-tune of the base model on curated data, for text-to-image and layout generation | Apache-2.0 | 2026-10-16 |
Select a model above for curated weight downloads, tutorials, and related information.
Capabilities
- Layout control: prompts can be a JSON document that places text and objects by bounding box, so composition and text placement are data you can edit
- Text rendering: legible English and Chinese text, with automatic line breaks when the prompt does not request them
- Long prompts: up to 4K tokens
- Prompt formats: natural language, booru-style tag sequences, or layout JSON
- Fine-tuning: the base checkpoint is published specifically for LoRA training, and the repository ships prompt-enhancer skills and a user guide for training and inference
Comments
Sign in with GitHub to join the discussion.