DreamLite unified image model hits sub-second mobile performance

DreamLite's unification of generation and editing in a sub-400M parameter model is a masterclass in efficiency for the mobile AI era.

• –Eliminates the need for separate models for generation and editing, drastically reducing memory footprint on RAM-constrained mobile devices
• –Step distillation enables high-quality 4-step inference, making "real-time" editing a reality without cloud latency or costs
• –Spatial "In-Context" conditioning allows the model to switch between tasks seamlessly based on input structure
• –Privacy is a core feature; local execution ensures sensitive image data never leaves the device's NPU
• –While the model is compact, its reliance on high-end NPUs marks a growing divide between flagship and budget mobile hardware capabilities