MOSS Image: The Apache 2.0 Text-to-Image Model With JSON Layout Control

OpenMoss released MOSS Image, a 5-billion-parameter text-to-image diffusion transformer trained from scratch and published under an Apache 2.0 license covering both weights and code. MOSS Image pairs its diffusion transformer with Qwen3.5-9B as the text encoder, accepting prompts up to roughly 4,000 tokens, and adds JSON layout control that places text and objects at specified bounding boxes before generation; on OpenMoss's own long-text benchmark the model scores close to Nano Banana and GPT Image, though independent third-party evals are not yet available. The Apache 2.0 terms distinguish MOSS Image from Qwen Image's non-commercial research license and the gated Flux community license, permitting commercial use, redistribution and modification, and community 8-step and 4-step LoRAs, quantized checkpoints and a full fine-tune appeared within a week of release.

MOSS Image: The Apache 2.0 Text-to-Image Model With JSON Layout Control MOSS Image is a 5B Apache 2.0 text-to-image model with strong text rendering and JSON layout control. Here's how to run it and its limits. What is MOSS Image? MOSS Image is a 5 billion parameter text-to-image diffusion transformer released by OpenMoss, the team behind MOSS TTS and the MOVA video model. It’s trained from scratch, released under an Apache 2.0 license for both weights and code, and it stands out for two things: unusually strong text rendering for its size, and a JSON-based layout system that lets you place text and objects at specific coordinates before generation even starts. It runs on Hugging Face, in ComfyUI, and through inference backends like SGLang and vLLM/Omni. TL;DR - MOSS Image is a 5B parameter diffusion transformer from OpenMoss, trained from scratch and released under a full Apache 2.0 license covering both weights and code. - The model uses Qwen3.5-9B as its text encoder, which means a 9B-parameter language model is reading your prompt up to roughly 4,000 tokens , a big upgrade over the small text encoders older image models relied on. - Its standout feature is JSON layout control , where you describe a scene, then list elements text or objects with bounding boxes, giving you precise placement that’s hard to get from prompting alone. - On the lab’s own long-text benchmark, MOSS Image scores close to Nano Banana and GPT Image, putting it well ahead of other open-weight models at text rendering, though independent third-party evals aren’t available yet. - It is text-to-image only , with no image-to-image or editing support, so injecting a specific person, object, or style requires training a LoRA rather than conditioning on a reference photo. - The Apache 2.0 license is the real differentiator versus models like Qwen Image or Flux, which carry non-commercial or community license restrictions that limit commercial use. - Community-made 8-step and 4-step LoRAs , quantized checkpoints, and a full fine-tune already exist within a week of release, and OpenMoss has said it’s working on an official few-step distillation. Everyone else built a construction worker. We built the contractor. One file at a time. UI, API, database, deploy. Why does the Apache 2.0 license matter? Most strong open-weight image models aren’t actually free to build on. Qwen Image ships under a non-commercial research license. Flux models are gated under the Flux community license, which restricts commercial deployment. MOSS Image is different: both the model weights and the training/inference code are Apache 2.0, which means no restrictions on commercial use, redistribution, or modification. You can fine-tune it, sell access to it, bundle it into a product, or build a service on top of it without a licensing conversation. That openness extends to content generation too. The model can produce both safe-for-work and NSFW output, since there’s no built-in content gate baked into the license terms the way there is with many hosted or community-licensed models. How does the text encoder affect prompt quality? MOSS Image pairs its 5B diffusion transformer with Qwen3.5-9B as the text encoder, accepting prompts up to around 4,000 tokens. That’s a meaningfully larger and more capable language model than what older text-to-image systems used for prompt encoding. Smaller text encoders are a major reason earlier diffusion models struggled with long, complex, or multi-part instructions, they simply couldn’t represent that much nuance. Having a 9B-parameter LLM read the prompt gives MOSS Image better handling of long captions, multi-element scenes, and the kind of structured instructions needed for layout-based generation. It’s also why the model can follow stylistic instructions watercolor, anime, specific artists and generate convincing lighting effects and bokeh backgrounds without needing hand-crafted prompt tricks. How does JSON layout control work? Instead of relying purely on natural language to describe where things go, MOSS Image accepts a JSON structure alongside the text prompt. You describe the overall scene, then list individual elements text strings or objects each paired with a bounding box defining where it should appear in the frame. The model was trained to respect this layout directly, so it’s not a post-processing hack or an inpainting workaround bolted onto a generic model. In practice this solves two recurring pain points with text-to-image generation: - Text placement accuracy. Asking for multiple lines of text a title, a date, a venue in a single unguided prompt often results in garbled or misplaced text as complexity increases. Defining a bounding box for each text element and specifying its content inside that box produces noticeably more reliable results. - Iterative control. If you generate an image and decide an element needs to move or change, you can adjust the layout JSON and regenerate rather than starting over or fighting with inpainting masks. Because writing JSON by hand is tedious, OpenMoss provides a layout prompt enhancer, and the same pattern works well as a two-stage pipeline: have an LLM turn a plain-language brief into prompt-plus-JSON, then feed that into MOSS Image for generation. What can MOSS Image actually generate well? At default settings 50 sampling steps, classifier-free guidance of 3, using the SFT checkpoint , MOSS Image handles: - Photorealistic portraits , with skin texture and fine detail holding up under scrutiny. - Product shots , including clean backgrounds and realistic bokeh. - Lighting , both for photorealistic scenes and stylized or artistic renders like watercolor. - Style transfer , including anime and specific artistic styles. - Text rendering , in both English and Chinese, including stylized contexts like chalkboard menus or multi-step infographics. - Flexible aspect ratios , from about 0.25 to 4 megapixels, with 1 to 2 megapixels generally producing the best quality. Other agents ship a demo. Remy ships an app. Real backend. Real database. Real auth. Real plumbing. Remy has it all. On the lab’s internal long-text benchmark, MOSS Image scores just behind Nano Banana and only a few points behind GPT Image 2, putting it well ahead of other open-weight models at rendering accurate text. That benchmark is self-reported, so treat it as a strong signal rather than an independently verified result. Text accuracy also degrades when you push too many words into a single image, and larger closed models likely still edge it out on dense, text-heavy layouts. What can’t MOSS Image do? The biggest limitation is that MOSS Image is text-to-image only. There’s no image-to-image conditioning and no true editing pipeline. You can use standard ComfyUI inpainting techniques, but you can’t take a real photo and edit it the way you can with Nano Banana or Qwen Image Edit. This matters most if you’re trying to put a specific, real person into a generated scene. Since the model can’t condition on a reference image, it either needs to already know that person from training data it recognizes some celebrities but not others, less well-documented public figures may simply not generate accurately or you need to train a LoRA to inject them into the weights. That’s a meaningfully higher bar than dropping in a reference photo, though it’s also the main avenue for customization given the Apache 2.0 base model release. Is MOSS Image worth running locally? If your priority is text-heavy images posters, infographics, UI mockups with precise element placement, and you need a license that allows commercial use without restriction, MOSS Image is one of the more compelling open options available right now. The combination of a capable text encoder, trained-in layout control, and full Apache 2.0 licensing is uncommon among models in this size class. If you need to edit existing photos, swap faces into real images, or do anything that requires conditioning on an input image, this isn’t the right tool yet. You’d need to wait for OpenMoss or the community to release an editing-capable variant. Performance out of the box is also slow at the default 50-step setting, especially on older or smaller GPUs. The community has already responded: within about a week of release, people had published 8-step and 4-step LoRAs, int8 and quantized checkpoints, and at least one full fine-tune of the base model. OpenMoss has also said it’s working on an official few-step distillation, so a “turbo” variant is likely coming. Frequently Asked Questions What license does MOSS Image use? Apache 2.0, covering both the model weights and the training/inference code. This permits commercial use, modification, and redistribution without the restrictions found in non-commercial research licenses or community licenses used by comparable models. Can MOSS Image edit existing photos? No. MOSS Image is a pure text-to-image model with no image-to-image or editing capability. You can use generic inpainting tools within ComfyUI, but there’s no native conditioning on an input photo. How do I control where text appears in a MOSS Image generation? Use the model’s JSON layout feature. Define the overall scene, then list each text or object element with a bounding box specifying its position. This produces more reliable placement and spelling than relying on a plain text prompt alone, especially for multi-line or multi-element compositions. Does MOSS Image know specific real people? Seven tools to build an app. Or just Remy. Editor, preview, AI agents, deploy — all in one tab. Nothing to install. It depends on how well-represented that person is in the training data. Widely photographed celebrities often generate recognizably, while less common public figures may not. Since there’s no image conditioning, getting an accurate likeness for an underrepresented person requires training a custom LoRA. What hardware or tools do I need to run MOSS Image? It runs natively in ComfyUI and is also supported through SGLang and vLLM/Omni. The official release includes a base model for fine-tuning and LoRA training and an SFT model for direct generation . Community quantized checkpoints and step-distilled LoRAs are already available for faster inference on more modest hardware.