MOSS-Image: Open Weights, Layout Control and ComfyUI Support

OpenMOSS released MOSS-Image on October 16, 2026, giving developers downloadable image-generation weights with structured control over where objects and text appear. The release includes a base checkpoint and a supervised fine-tuned version, both under Apache-2.0. ComfyUI support was merged the same day.

For teams building poster generators, illustrated explainers or local creative tools, the appeal is specific: a layout can be described in data before the image is generated. There are still practical limits around sampling speed, text accuracy and the memory needed for the complete pipeline.

What is MOSS-Image?

The official project page describes a text-to-image family with a 5B backbone and prompts up to 4K tokens. It accepts ordinary descriptions as well as structured layouts, and shows examples involving Chinese and English text, illustration and photography.

MOSS-Image-Base is the starting checkpoint for further adaptation. MOSS-Image-SFT adds fine-tuning on curated data. The team's whiteboard example shows clearer text grouping with SFT, but that example alone doesn't establish which version will work better across a whole design workflow.

The SFT model card publishes the weights and bundled inference code under Apache-2.0. The project page notes that third-party components retain their respective licenses. Teams packaging the pipeline should retain and review those component notices too.

How JSON layout control works

A layout prompt describes the overall scene, then lists individual text and object elements with bounding boxes. A poster, for example, can put a title in one region and an illustration beneath it. The developer supplies the intended wording and geometry together.

The user guide labels layout prompts experimental. Its boxes use pixel coordinates measured from the canvas centre, with coordinates ordered as minimum y, minimum x, maximum y and maximum x. The requested image dimensions need to match the layout's coordinate system.

That gives an application a useful way to translate a template into a generation request. It still needs to inspect the output. Check spelling and line breaks, then check whether the generated elements actually respect their boxes. The guide warns that automatic text layout may differ from expectations and suggests explicit line breaks when they matter.

Changing the layout prompt generates a revised composition. Developers needing reference-image editing should evaluate that separately: the guide discusses reference-based generation as further post-training work.

ComfyUI support and the 5B memory question

The ComfyUI integration merged on October 16. Its description identifies a 5B diffusion transformer using a Qwen3.5-9B text encoder and the FLUX.2 latent space. The headline parameter count therefore describes only part of the inference pipeline.

Comfy-Org's model package separates diffusion-model files, text encoders and the VAE, and includes quantized options. Use a ComfyUI version containing the new support and follow the package's file-placement instructions. An existing installation may need updating before it recognizes the model.

The official Python quick-start targets an NVIDIA GPU with CUDA. OpenMOSS also provides CPU-offload modes, trading memory use against inference speed. Hardware planning should account for the encoder and VAE alongside the diffusion model.

Speed, cost and early adoption

The recommended base settings use at least 50 inference steps. That is worth considering when comparing it with distilled models designed for fewer steps. A step count alone cannot predict end-to-end latency across different architectures and hardware.

There is no published per-image API price in the release materials reviewed here. The SFT model card lists no Hugging Face inference provider at the time of checking. Downloadable weights leave the cost of hardware, hosting and operation with the team running them.

Early community discussion shows interest in local generation and fine-tuning, alongside reports of slow generation and inconsistent results. Those posts are useful for identifying tests to run, rather than establishing a quality ranking.

A sensible evaluation starts with a few real layouts: dense labels, awkward line breaks and an off-centre object. Run both checkpoints, keep the failures and count the outputs that need manual repair. That will show whether structured prompting saves work for the application you are building.

I'm Paras. I work at the intersection of AI, engineering and distribution.

I help technical founders turn model launches and product updates into clear stories that developers and founders can understand and share.

Book a 30-minute call.