PRISM: Feed-Forward Single-Image 3D Reconstruction via Geometric Warp-Residual Modeling

2026-06-24 · Source: Takara TLDR - Daily AI Papers · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Robotics & Autonomous Systems, Emerging Technologies & Innovation · Depth: Expert, quick

Summary

PRISM is a novel feed-forward framework designed for single-image 3D scene reconstruction, addressing a fundamental challenge in computer vision. It overcomes the practical deployment limitations of iterative diffusion sampling inherent in existing camera-controlled video diffusion models. PRISM achieves this by decomposing multi-view latent prediction into a parameter-free geometric prior and a learned residual correction, eliminating the need for diffusion sampling during inference. The framework employs a two-stage training strategy, combining latent supervised distillation for geometric generalization and perceptual fine-tuning for appearance quality optimization. Extensive experiments on three benchmarks demonstrate PRISM delivers competitive reconstruction quality while dramatically reducing inference time to only 36 seconds per scene.

Key takeaway

For Computer Vision Engineers developing real-time 3D reconstruction applications, PRISM offers a compelling alternative to diffusion-based methods. Its feed-forward architecture dramatically reduces inference time to 36 seconds per scene, making it suitable for deployment-constrained environments. You should evaluate PRISM for projects requiring rapid single-image 3D scene generation without significant quality compromise.

Key insights

PRISM enables fast, feed-forward single-image 3D reconstruction by correcting geometric warps with a learned residual.

Principles

• Geometric forward warping covers the majority of target view data
• Decomposition into a prior and residual improves efficiency
• Two-stage training aids generalization from synthetic data

Method

Decompose multi-view latent prediction into a parameter-free geometric prior and a learned residual correction. Train in two stages: latent supervised distillation and perceptual fine-tuning.

In practice

• Apply geometric warping as a strong initial prior
• Use residual learning for fine-grained corrections
• Employ two-stage training for synthetic data generalization

Topics

• Single-Image 3D Reconstruction
• Feed-Forward Networks
• Geometric Warping
• Residual Learning
• Multi-view Latent Prediction
• Computer Vision

Best for: Research Scientist, AI Scientist, Computer Vision Engineer, Machine Learning Engineer

Related on AIssential

• Feed-Forward 3D Scene Modeling: A Problem-Driven Perspective — Feed-forward 3D reconstruction offers efficient, generalizable scene modeling by directly mapping 2D inputs to 3D representations.
• PRISM: Synergizing Vision Foundation Models via Self-organized Expert Specialization — PRISM synergizes Vision Foundation Models through a dual-stream Mixture-of-Experts framework with self-organized expert specialization.
• 3D Consistency Optimization for Self-Supervised Monocular Video Depth Estimation — Recasting monocular video depth as multi-view 3D reconstruction with 3D consistency optimization improves geometric accuracy and coherence.
• MoRe: Motion-aware Feed-forward 4D Reconstruction Transformer — MoRe efficiently reconstructs dynamic 4D scenes from monocular video by disentangling motion from static structure.
• JointEdit3D: Feed-Forward 3D Scene Editing in a Unified Latent Space — JointEdit3D enables feed-forward 3D scene editing by unifying RGB-geometry reconstruction and generation in a single latent space.
• Latent Spatial Memory for Video World Models — Latent spatial memory in diffusion models significantly boosts video generation speed and reduces memory by avoiding pixel-space reconstruction.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Takara TLDR - Daily AI Papers.