ByteDance's DreamLite runs image generation and editing offline on phones

Waiting several seconds for a cloud server to erase an unwanted object from your vacation photo is annoying enough, but losing connection on a flight or in the subway turns your mobile AI tools completely useless.

ByteDance introduced DreamLite, an ultra-compact 0.39-billion-parameter diffusion model engineered specifically to run entirely offline on smartphones. Unlike most lightweight attempts that only handle basic text-to-image prompts, DreamLite combines both text-to-image generation and text-guided image editing into a single compact architecture.

On an iPhone 17 Pro, it churns out a 1024×1024 portrait of a young woman with flowers in roughly 3 seconds. When tested on a Xiaomi 14 with 4-step distillation, both generation and editing tasks drop under 1 second.

Pocket-sized diffusion with zero latency

Most existing on-device diffusion networks are one-trick ponies focused solely on generating images from scratch. Once you want to alter a specific object or tweak the style, you typically have to call home to massive cloud models.

DreamLite bridges that gap. Beyond standalone generation, users can feed it existing photos and modify them using plain text. You can turn a regular snapshot into an oil painting, transform a sunny landscape into a frozen winter lake, swap out the background behind a cactus, remove a bird from the frame, or overlay a cyberpunk aesthetic.

Because the model is only 390 million parameters, it naturally cannot match the pore-level skin texture or hyper-detailed animal fur of massive server-side models. But when stacked against models of similar or even slightly larger sizes, its visual fidelity and speed stand out noticeably.

How they shrunk two tasks into one network

Packing generation and instruction-based editing into a sub-billion model usually requires bulky adapter networks or separate backbones, which blow past mobile memory budgets.

ByteDance solved this by building DreamLite on a pruned mobile U-Net. Instead of complex routing, the model simply concatenates latents horizontally: it takes a target and a blank canvas for text-to-image generation, and a target alongside the source image for editing tasks.

To ensure the miniature network could handle both workflows without destabilizing, the team used a task-progressive joint pretraining regimen, moving from text-to-image to editing before training jointly, followed by supervised fine-tuning and reinforcement learning. On standard benchmarks, it achieved 0.72 on GenEval and 4.11 on ImgEdit, outclassing other on-device models and matching several server setups.

Why it matters

For everyday smartphone users, local execution means zero waiting for network queues and zero privacy exposure. Editing sensitive family pictures no longer requires sending personal data across remote servers, and you can edit vacation shots offline mid-flight.

For hardware makers and the broader industry, DreamLite marks a shift from experimental toy models to usable local features. Achieving sub-second latency on a 1024×1024 canvas with just 4 denoising steps proves that mobile processors can handle complex creative tasks without burning battery or hogging system RAM.

We think ByteDance made a smart trade-off here. Instead of chasing photorealistic perfection at the cost of tens of billions of parameters, DreamLite zeroes in on what actually matters on a handheld device: tiny storage footprint, low latency, and versatile everyday editing. For on-device computing, instant utility beats raw parameter count every time.