High-fidelity image generation often struggles when images must contain dense, readable text or tightly constrained object layouts. This model targets that gap by combining a compact 5B single-stream diffusion transformer with native JSON layout prompts and specialized prompt-enhancement tooling, making it easier to produce images with explicit spatial composition and legible long-form text in English and Chinese.
Key Capabilities
- Native layout control: accept structured JSON layouts (summary.caption + items) with centralized absolute bounding boxes (yxyx) so you can specify object positions and rendered text regions directly.
- Strong bilingual text rendering: trained and evaluated for long-form English and Chinese text, designed to improve text coverage in image outputs compared with typical open-weight generators.
- Practical inference pipeline: distributed as a custom Diffusers pipeline (MOSSImagePipeline) with recommended defaults (Flow-matching Euler sampler, 50 steps, CFG≈3, negative-branch layer skip at index 22). Uses bfloat16 for transformer/text encoder and float32 for VAE and initial noise; bundled VAE uses the FLUX.2 small decoder.
- Long prompts and agent tooling: supports prompts up to ~4K tokens and includes a Prompt Enhancer skill to convert ideas or pixel drafts into compact layout JSON for controlled generation.
- Open license and ecosystem integrations: model weights and bundled inference code released under Apache‑2.0 and designed for PyTorch/Diffusers; community integrations include a Gradio demo and ComfyUI plugin.
Who it's for and tradeoffs
Great fit if you need images that contain readable, multi-line text or require explicit spatial layout (e.g., posters, whiteboards, UI mockups, annotated graphics) and prefer open-weight tooling you can run or fine-tune. Also useful for developers who want programmatic layout-to-image workflows via JSON.
Look elsewhere if you need massively large backbones for ultra-high visual fidelity at extreme resolutions without additional upscaling, or if you require turnkey hosted inference with vendor SLAs; the model is GPU‑centric (CUDA + PyTorch) and inference uses multiple transformer forwards per step, so production costs can be nontrivial. The SFT checkpoint in the family can yield clearer text grouping for some tasks.
