Open-source tools meant for editing photos are accidentally better at understanding 3D space than many specialized vision models.
Image-editing models perform complex tasks like depth estimation and semantic segmentation with zero specific training. These models were built to change images, but the process of learning to edit forced them to master the underlying geometry and meaning of a scene. They effectively learned how the world is put together as a byproduct of learning how to alter it. This discovery allows practitioners to use general-purpose editing tools for high-precision vision tasks without any fine-tuning. It shows that creative generative tasks are a backdoor to deep spatial understanding in artificial intelligence.
Open-Source Image Editing Models Are Zero-Shot Vision Learners
arXiv · 2605.04566
Recent studies have shown that large generative models can solve vision tasks they were not explicitly trained for. However, existing evidence relies on closed-source models~(Veo~3, Nano Banana Pro) or requires task-specific instruction tuning, leaving open whether publicly available image-editing models possess zero-shot vision abilities out of the box.We conduct a systematic evaluation of three open-source image-editing models -- Qwen-Image-Edit, FireRed-Image-Edit, and LongCat-Image-Edit -- o