With MOSS-Image, text and objects are placed directly into the composition

OpenMOSS releases MOSS-Image, an open-weight image generation model that enables text and object placement through JSON instructions.

Instead of asking an AI model to place a title approximately above an illustration, what if users could specify the coordinates of each element directly? With MOSS-Image, OpenMOSS introduces an image generation family designed to combine written instructions with structured spatial control. Available in Base and SFT versions, the models accept natural-language descriptions, tag sequences, and layouts defined in JSON. The project focuses particularly on controlling where text and objects appear, targeting applications such as posters, visual documents, diagrams, and graphic compositions. Image Generation With Layout Awareness

Most image generators allow users to describe a composition through natural-language prompts.

A user might request a poster with a centered headline, a character on the left, and an illustration near the bottom.

However, these instructions remain subject to the model's interpretation. Elements may appear in unexpected positions, and spatial relationships are not always followed accurately.

MOSS-Image introduces a different method: describing individual visual elements and specifying their positions within a structured data format.

The models accept three main input formats: natural-language prompts, tag sequences, and JSON layouts.

In the third case, users define the overall image context separately from the objects that should appear.

Each element can have its own description, type, and placement region.

This does not replace conventional creative prompting. It adds a framework intended to provide more control over composition. Positioning Text and Objects With Coordinates

Spatial control relies on rectangular regions known as bounding boxes.

Users can define a poster by specifying that a headline should occupy the upper area, an illustration should appear in the center, and secondary text should be placed elsewhere.

These instructions are stored in a JSON structure containing fields such as `summary` and `items`.

The first describes the overall context, scene characteristics, and desired visual style.

The second lists individual elements, including their descriptions and coordinates.

The layout format uses positions relative to the center of the image, with boundaries defining each region.

In a demonstration published by OpenMOSS, the model generates a minimalist travel poster with the title EXPLORE near the top and a blue mountain illustration underneath.

Users can modify the headline, move the illustration region, or change an object's description directly in the layout data.

This approach brings image generation closer to certain principles used in graphic layout tools, where individual elements have defined positions.

The model still interprets the instructions generatively, however. Coordinates guide the result rather than guaranteeing perfectly exact placement in every generation. A Focus on Text Rendering

MOSS-Image also emphasizes generating readable text and organizing written content within images.

This capability is important for posters, presentations, explanatory graphics, and other visual materials combining design and information.

Image generation models can produce distorted characters, change words, or struggle when multiple blocks of text appear in the same composition.

OpenMOSS presents MOSS-Image as a system designed to improve accuracy, particularly for English and Chinese text.

Its demonstrations include graphic compositions, annotations, and handwritten elements.

One example shows a handwritten Chinese note attached to a refrigerator door, with several lines of text integrated into the scene.

JSON control also separates the exact text to be displayed from instructions describing its appearance and placement.

These examples illustrate the model's capabilities but do not guarantee flawless reproduction of every string, especially in lengthy or densely arranged compositions. Base and SFT: Two Versions for Different Workflows

The MOSS-Image family includes two checkpoints that share the same architecture and generation interface.

MOSS-Image-Base is the foundational model, intended for research, further training, and adaptation to specialized tasks.

MOSS-Image-SFT is a supervised fine-tuned version developed from the base checkpoint using curated data.

The additional training is intended to improve the quality of immediately usable generations.

OpenMOSS provides a comparison involving a whiteboard-style composition.

In that example, SFT produces more readable text grouping and a more organized visual structure.

Both versions support natural-language prompts, tags, and JSON layouts.

The choice therefore depends on the intended workflow: experimentation and customization with Base, or generation with additional fine-tuning through SFT. A Five-Billion-Parameter Generative Backbone

ModelScope describes MOSS-Image-Base as an 8.56-billion-parameter model, while OpenMOSS's official documentation identifies a generative backbone containing approximately five billion parameters.

These figures do not necessarily describe the same architectural components.

The project uses a Diffusion Transformer to generate visual content.

It also includes separate components for interpreting instructions and encoding or decoding images.

The technical documentation references a text encoder from the Qwen family and a FLUX.2 visual decoder.

The difference between ModelScope's reported figure and the backbone size should therefore be distinguished rather than presented as conflicting measurements of the same component.

OpenMOSS also reports training across heterogeneous GPU platforms.

The models support prompts of up to 4,000 tokens, allowing relatively detailed scene descriptions and more complex layout specifications. From Photorealism to Illustration

MOSS-Image is not limited to posters featuring typography.

OpenMOSS's examples also demonstrate illustrations, photographic-style images, and anime-inspired compositions.

Users can rely on conventional text prompting when they do not need to define every element