MOSS-Image-Base: 5B Image Model That Controls Text Placement Down to the JSON Layout

MOSS-Image-Base is a 5B-parameter text-to-image base model released by the OpenMOSS team. It accepts natural language prompts, tag sequences, and JSON layouts as input, enabling precise control over text placement and object arrangement. Distributed under the Apache 2.0 license, it can be freely used in commercial environments.

Unlike existing generative models, passing a JSON-structured layout as a prompt allows for explicit specification of text and object coordinates within a scene. This results in clearer text grouping and layout alignment when generating structured images such as whiteboards or sticky notes. Compared to the SFT version, the Base model provides the foundational infrastructure for this structural control.

It works with the Diffusers library and uses the FLUX.2 small VAE. Inference is performed in bfloat16 precision on NVIDIA GPUs in a CUDA environment. It also supports a native ComfyUI plugin for workflow integration.

It is suitable for graphic design, infographic generation, or UI mockup creation where precise text rendering and object position control are required. It can also be used to build pipelines that convert ideas into detailed captions or spatial layouts when combined with GPT-6 family agents.