ByteVisionLab/DreamLite

What it solves

DreamLite is a compact diffusion model designed for high-resolution image generation and editing directly on mobile devices. It eliminates the need for cloud dependency by providing a lightweight architecture that can generate or edit 1024x1024 images in seconds on modern smartphones.

How it works

The project uses a pruned mobile U-Net backbone (0.39B parameters) and unifies text-to-image and image-editing tasks through In-Context Spatial Concatenation in the latent space. To achieve high speed, it employs progressive step distillation, allowing the "Mobile" variant to perform inference in just 4 steps. It is compatible with the Hugging Face Diffusers library and supports 4-bit quantization for on-device deployment.

Who it’s for

It is intended for developers and researchers building on-device AI applications that require fast, high-resolution image creation and editing without relying on external servers.

Highlights

• Unified Model: A single network architecture handles both text-to-image generation and text-guided image editing.
• Ultra-Fast Inference: Capable of producing 1024x1024 images in approximately 3 seconds on an iPhone 17 Pro.
• On-Device Privacy: Operates fully locally with zero cloud dependency.
• Customizable: Includes support for LoRA fine-tuning for domain-specific datasets.
• Deployment Ready: Provides a complete iOS deployment reference including CoreML and mlx-vlm quantization scripts.