Emu Edit

Experience state-of-the-art instruction-based image editing with Emu Edit, equipped for a broad range of editing and computer vision tasks.

Image Editing Updated 7 minutes ago

Visit Website

Emu Edit's Top Features

Multi-task image editing

Region-based editing

Free-form editing

Computer vision tasks: detection and segmentation

Learned task embeddings

Few-shot learning

Task inversion

Benchmark with seven tasks

State-of-the-art performance

Unprecedented task diversity

Frequently asked questions about Emu Edit

Emu Edit is a multi-task image editing model designed for instruction-based editing, supporting tasks like region-based editing, free-form editing, and computer vision tasks such as detection and segmentation.

Emu Edit achieves multi-task learning by adapting its architecture to handle multiple tasks and training on a wide variety of editing and computer vision tasks.

Learned task embeddings in Emu Edit steer the generation process toward the correct generative task, enhancing the model's ability to execute editing instructions accurately.

Yes, Emu Edit can adapt to new tasks using few-shot learning where it updates a task embedding to fit the new task, even with limited labeled examples.

Emu Edit can perform a range of tasks including region-based editing, free-form editing, computer vision tasks like detection and segmentation, and additional tasks like super-resolution and contour detection.

Task inversion in Emu Edit keeps the model weights frozen and updates a task embedding to swiftly adapt to new tasks, making it efficient for scenarios with limited labeled examples.

A benchmark including seven different image editing tasks such as background alteration, global changes, style alteration, object removal, object addition, localized modifications, and texture changes is released with Emu Edit.

Emu Edit was developed by a team of researchers including Shelly Sheynin, Adam Polyak, Uriel Singer, Yuval Kirstain, Amit Zohar, Oron Ashual, Devi Parikh, and Yaniv Taigman.

Emu Edit handles free-form editing by treating it as a generative task, leveraging its trained model and learned task embeddings to generate accurate edits based on instructions.

You can access the Emu Edit model and its benchmark datasets from the official website where downloads for both are available.