Align3D: Image-Structure Co-Aligned 3D Scene Generation with Observability-Aware Fidelity-Preserving Guidance

Align3D is an image‑conditioned diffusion model that generates 3D indoor layouts faithful to a single input photograph while remaining physically plausible beyond the observed view. It infers a confidence‑calibrated scene graph from the photo and trains a set‑diffusion layout denoiser with image‑structure co‑alignment, balancing image and scene‑graph alignment through confidence‑weighted fusion. The model employs observability‑aware fidelity‑preserving guidance to reconcile conflicting scene‑quality gradients, using an object‑level confidence mask so under‑observed degrees of freedom can incorporate physical priors without disrupting the image‑graph match. On the 3D‑FRONT dataset, Align3D outperforms recent baselines in image‑alignment and scene‑quality metrics, and can map unseen photographs to layouts in a single feed‑forward pass.

Machine-generated by The Flow from the publisher's headline and feed description — not written or checked by a human. The full article lives at arXiv Computer Vision.

• diffusion
• safety