OpenMOSS drops MOSS-Image, a 5B image model with sharp text and layout control

MOSS-Image is a new open-source, 5-billion-parameter text-to-image model from OpenMOSS that packs precise spatial layout control and clean text rendering into a surprisingly small footprint.

Everything is licensed under Apache 2.0 with weights and code fully public, meaning you can freely modify it or plug it into commercial production pipelines without legal headaches.

What it can do

Most open diffusion models still struggle when you ask them to put specific objects in exact spots or print legible text. MOSS-Image tackles both problems head-on.

Its headline feature is native bounding-box layout control. Similar to tools like Krea, you can sketch boxes directly on the canvas to dictate where each character, prop, or background piece belongs, and the model follows those boundaries faithfully.

It also handles text rendering cleanly. The model can accurately spell words inside scenes, from handwritten fridge notes in Chinese to crisp signage, and accepts long prompt inputs of up to 4K tokens. That makes it uniquely useful for comic panels, UI wireframes, infographics, and anime or game concept art.

YouTube creators have noted that it feels much less restricted out of the box than Flux 3. The main trade-off right now is flexibility: it is strictly a text-to-image model, meaning you cannot edit existing images through natural language prompts.

Specs and how to run it

The model is built on a compact 5B DiT backbone. Its core Transformer weights take up only around 10GB of storage, making it very accessible for standard consumer GPUs.

OpenMOSS released two distinct checkpoints. MOSS-Image-Base is the clean foundation intended for downstream training, while MOSS-Image-SFT has undergone supervised fine-tuning for much higher visual quality and aesthetic polish. If you just want to generate images, SFT is the one to grab.

Tooling support was ready on day one. ComfyUI supports MOSS-Image natively in its master branch without custom nodes, and it runs on inference engines like SGLang and vLLM-Omni. For fine-tuning, the companion KohakuMossTrainer supports LoRA and LyCORIS on GPUs with as little as 6GB of VRAM.

If you prefer not to manage local weights, you can also test it immediately via the playground on the MOSI API platform.

Why it matters

For comic creators, indie game artists, and solo designers, MOSS-Image fills a critical gap. Instead of juggling complex external ControlNet stacks just to keep characters from overlapping, you can sketch layout boxes natively and get readable text on the first pass.

We think OpenMOSS made the right call by capping the backbone at 5B parameters. Instead of chasing bloated parameter races that require enterprise clusters, they delivered a fast, lightweight tool that everyday developers can fine-tune on consumer hardware.

It won't replace your interactive inpainting workflows yet because it lacks native image-to-image chat editing. But as a scratchpad for structured layouts, posters, and sequential art, it is easily one of the most practical open models to drop this season.