# Emu Edit: Precise Image Editing via Recognition and Generation Tasks

**Authors:** Shelly Sheynin, Adam Polyak, Uriel Singer, Yuval Kirstain, Amit Zohar, Oron Ashual, Devi Parikh, Yaniv Taigman  
**arXiv:** [2311.10089](https://arxiv.org/abs/2311.10089)  
**Date:** 2023-11-16  
**Summary page:** https://www.summarizepaper.com/en/paper/2311.10089v1/

## Key points

- Emu Edit is a multi-task image editing model
- It aims to improve accuracy and performance of instruction-based image editing
- Trained to perform region-based editing, free-form editing, and Computer Vision tasks
- Formulated as generative tasks for precise edits based on natural language instructions
- Learns from task embeddings to enhance multi-task learning abilities
- Demonstrates outstanding performance in instruction-based image editing
- Generalizes well to new tasks with few labeled examples, even when high-quality samples are scarce
- Authors have released a challenging benchmark for assessing instructable image editing models like Emu Edit
- Includes seven different image editing tasks and comprehensive evaluation framework
- Achieves state-of-the-art results and demonstrates robustness in handling various editing tasks

## Summary

Emu Edit is a multi-task image editing model that aims to improve the accuracy and performance of instruction-based image editing. To address this challenge, Emu Edit is trained to perform a wide range of tasks, including region-based editing, free-form editing, and Computer Vision tasks. These tasks are formulated as generative tasks, allowing Emu Edit to generate precise edits based on natural language instructions. One key feature of Emu Edit is its ability to learn from task embeddings which guide the generation process towards the correct edit type and enhance the model's multi-task learning abilities. This contributes to Emu Edit's outstanding performance in instruction-based image editing. Furthermore, Emu Edit demonstrates the capability to generalize to new tasks with just a few labeled examples such as image inpainting, super-resolution, and compositions of editing tasks even when high-quality samples are scarce. To facilitate a more rigorous assessment of instructable image editing models like Emu Edit, the authors have released a challenging and versatile benchmark which includes seven different image editing tasks and provides a comprehensive evaluation framework for future research in this field. Overall, Emu Edit presents promising advancements in instruction-based image editing by achieving state-of-the-art results and demonstrating robustness in handling various editing tasks. Its multi-task learning capabilities and generalization abilities make it an effective tool for precise image editing using natural language instructions.