BlazeEdit: 195M Parameter Image Editing Model for Mobile Devices
Researchers have introduced BlazeEdit, a compact image-to-image diffusion model with only 195 million parameters, designed for on-device deployment on mobile devices. Unlike larger text-to-image models that require server-side inference, BlazeEdit eliminates text-conditioning components and uses a multi-task architecture to perform object removal, outpainting, tone correction, relighting, and sticker generation. The model addresses privacy risks and computational costs associated with server-based processing. The paper is available on arXiv under ID 2605.28067.
Key facts
- BlazeEdit is an image-to-image diffusion model with 195 million parameters.
- It is designed for on-device deployment on mobile devices.
- The model eliminates text-conditioning components.
- It consolidates object removal, outpainting, tone correction, relighting, and sticker generation.
- The paper is published on arXiv with ID 2605.28067.
- The model addresses privacy risks and computational costs of server-side inference.
- Existing text-to-image models for mobile range from 0.5B to 1B parameters.
- BlazeEdit is described as a generalist image editing model.
Entities
Institutions
- arXiv