MOSS–ImageBilingual text-to-image
Chinese and English text rendering appears in the project’s examples; the user guide recommends English prompts and says Chinese and other languages are also supported. Always inspect the final lettering.
More intention.
Less guesswork.
A creative workflow built around your words and your composition. Here is what the open model offers today, and what we hope to build around it.
The model is available. Our hosted product is not. “At launch” describes proposed initial scope; “Planned” describes longer-term ideas. Neither is a release commitment.
Create
Hosted generation & aspect ratios
Proposed prompt-to-image workspace with canvas sizes, a credit allowance, and downloadable results.
Conversational iteration
Explore revisions through a chat-like workflow, with version history to make changes understandable.
Motion & video
Research direction for future workflows; no video model or hosted video service is offered here.
Control
JSON layout prompts
Describe text and object regions using the project’s experimental structured input format.
Visual layout editor
Proposed interface for arranging regions and exporting a layout prompt. The home preview is only an illustration.
References, moodboards & personalization
Style and subject reference workflows, curated visual directions, and saved preferences for consistent work.
Long-form art direction
Published support for prompts up to roughly 4K tokens, with room for explicit copy and composition.
Edit
Upload & multi-turn editing
Bring an image and describe changes. This is a service concept, not a capability claimed for the released checkpoint.
Inpaint & outpaint
Selected-area edits and canvas expansion, subject to further model and workflow evaluation.
Upscale & retexture
Higher-resolution finishing and alternative surface treatments. No quality or resolution guarantee yet.
Style changes & transparency
Explore restyling and transparent exports, with actual support to be validated before launch.
Workflow
Base, SFT & ComfyUI
Download the checkpoints and use the official project’s custom-node plugin. Follow current upstream instructions.
Saved projects & credit tracking
Proposed essentials for the first hosted experience. Launch timing and limits are not yet set.
API, teams & brand kits
Potential programmatic access, shared workspaces, and LoRA hosting for team workflows.
Explore & model versions
A community discovery surface and explicit model selection. No public gallery of model outputs is live here.
What makes MOSS-Image different
Text meets structure.
Bilingual text rendering, JSON-based composition, and long prompts are central documented interests of this release. None implies guaranteed character accuracy.
A starting point you can study.
Apache-2.0 model weights and the adaptation-oriented Base checkpoint let you inspect and experiment with your own workflow. SFT offers another ready-to-evaluate starting point.
A workflow you can connect.
The official ComfyUI plugin provides a documented local path. You can explore the model before any service on this site launches.
Sources: MOSS-Image repository, Base model card, ComfyUI plugin instructions.
MOSS-Image vs typical
closed generators
Compare operating models, not an untested quality ranking. “Closed generators” describes hosted services such as Midjourney and ChatGPT Images; individual capabilities and terms vary.
| Consideration | MOSS-Image open model | Typical closed generator |
|---|---|---|
| Access | Downloadable Base and SFT weights | Hosted service access; model downloads are not the standard workflow |
| Composition | Experimental JSON regions plus text prompts | Prompt, reference, and editing controls vary by service |
| Bilingual text | Documented Chinese and English examples | Text support varies; assess your own brief |
| Customization | Adapt Base under its license; manage your own compute | Service-specific references, personalization, or editing tools |
| Workflow today | Local inference and official project ComfyUI plugin | Managed web or app interface |
| This site’s service | Waitlist only; no generation API yet | Consult each provider for current availability and terms |
Learning from Midjourney
Its documented workflow spans style references, personalization, moodboards, an editor, upscaling, retexturing, video, and Explore. Character and Omni reference naming depends on model version; newer Edit Model documentation describes a replacement reference workflow.
Learning from ChatGPT Images
Its help article describes creating images from natural-language prompts, editing uploaded or generated images by describing changes, transparent-background requests, and aspect-ratio choices. Asking for a style change is a conversational workflow, not a separate feature we verified.
Our point of focus
We are exploring a smaller experience around bilingual posters and explicit composition. Roadmap labels describe that ambition; they do not claim feature parity with established products.
Comparison sources, checked October 18, 2026: Style Reference, Moodboards, Personalization, Editor, Edit Model, Versions, Website & Explore, Video, Upscalers, ChatGPT Images. See the full reference list.