DreamLite: A 0.39B Mobile Model Matching Z‑Image for Real‑Time Text‑to‑Image Generation and Editing

DreamLite is a newly open‑sourced compact unified diffusion model (0.39 B parameters) from ByteDance that can generate images from text and edit images via textual prompts directly on a mobile device.

The architecture builds on a pruned mobile U‑Net backbone and introduces context‑spatial connections in the latent space to unify generation and editing. Training follows a task‑progressive joint pre‑training strategy (T2I → editing → joint) and employs Qwen3‑VL as the text‑embedding model; step‑distillation reduces inference to four steps.

On an iPhone 17 Pro, a 4‑bit Qwen‑VL encoder, fp16 Tiny VAE, and the UNet backbone produce or edit a 1024×1024 image in roughly three seconds, making DreamLite the first unified on‑device model that eliminates the need for separate generation and editing networks.

Benchmark results show DreamLite matches the quality of open‑source models such as Flux, Z‑Image and LongCat‑Image. The release provides two model variants that let users choose the optimal balance between visual realism and on‑device inference latency.

Model weights are currently undergoing safety review. Developers interested in early access can request it via email ([email protected]) with name, affiliation, and intended use, and must agree to ethical guidelines that forbid any illegal, violent, or harmful content.