The open-weight architecture combines a 7.1B parameter DiT with Qwen3-VL-8B to deliver a unified tool for image creation and modification. It handles RGBA layers to allow seamless compositing and accepts up to 10 reference images for precise local control over portraits and product textures. The system includes prompt enhancement LLMs and integrates with the Diffusers library.

A strict non-commercial license governs the model's use, which bars revenue generation from outputs including monetized videos. The tool features day-0 support in vLLM-Omni, SGLang-Diffusion, ComfyUI, and the Ostris AI Toolkit. Performance benchmarks on an RTX 4090 24GB GPU show 1024x1024 image generation in 18.7s with 22.7 GiB peak GPU memory, while training with alpha datasets can be completed in 750 steps.