G1NDiff: Fast 3D scene editing with reprojection-conditioned GAN-diffusion training on 2D datasets

Abstract

We present G1NDiff, a novel pipeline for fast 3D scene editing that maintains multi-view consistency and semantic alignment with text prompts. We design and train a one-step diffusion model G1NDiff including architectural modifications and a reprojection aware conditioning to preserve consistency across successive edits by generating new occluded regions in a single forward pass. To do this, our approach introduces an occlusion-aware on-the-fly reprojection algorithm to generate paired conditioning–target views from single 2D images, enabling the use of massive 2D image datasets for reprojection-conditioned training. We further propose a multi-stage dataset editing procedure that iteratively propagates edits across all views, enforcing geometric consistency by thresholding visible content and filling occluded regions. In experiments, our method achieves high CLIP alignment with textual edits and state of the art performance in cross-view consistency (measured by the MEt3R metric), while drastically reducing editing time.