AGAP: A Simple Technique for Editing 3D Neural Radiance Fields Explained in Plain Language

Every year Ant publishes hundreds of papers at top conferences and journals, many of which are open‑sourced or deployed in Alipay and partner scenarios. Starting today, Ant launches a "Plain Talk Tech" video column to introduce selected papers in an easy‑to‑understand format.

The featured technology, AGAP, is an academic result from Ant Technology Research Institute's Interactive Intelligence Lab that was open‑sourced on GitHub earlier this year. AGAP addresses the difficulty of editing 3D neural radiance fields by replacing the color field with a clear 2D appearance aggregate and using a projection field to map 3D points to 2D pixels for texture lookup, supporting multiple 3D editing modes efficiently and universally.

In lay terms, AGAP makes 3D image editing as simple as Photoshop for 2D photos: it first reduces a 3D scene to a 2D image, allows standard image editing, and then restores the edited content back to 3D without any loss of pixel detail or sharpness.

Ant researcher Wang Qiuyu, a post‑90s scientist, provides a full "translation" of AGAP’s technical principles and explains why 3D image editing has been hard to productize.

To see the technology in action, watch the four‑minute video linked below.