Uni-Edit: Intelligent Editing Is A General Task For Unified Model Tuning

Researchers introduced Uni-Edit, a general intelligent image-editing task for tuning unified multimodal models.

An automated synthesis pipeline converts VQA data into complex editing instructions, creating the Uni-Edit-148k dataset.

Training on Uni-Edit-148k improved understanding, generation, and editing performance without auxiliary training stages.

The results suggest one well-designed editing task can replace more complex mixed multi-task training.