Ouroboros: Single-step Diffusion Models for Cycle-consistent Forward and Inverse Rendering

Shanlin Sun*1, Yifan Wang*2, Hanwen Zhang*3, Yifeng Xiong1, Qin Ren2, Ruogu Fang†4, Xiaohui Xie†1, Chenyu You†2
1University of California, Irvine    2Stony Brook University    3Huazhong University of Science and Technology
4University of Florida
Paper Code
Teaser image.

Single-step diffusion models for forward and inverse rendering in cycle consistency. Ouroboros decomposes RGB inputs into intrinsic maps and reconstructs images through a mutually reinforced inverse-and-forward rendering loop. It also extends to temporally consistent video inverse rendering without dedicated finetuning.

Abstract

While multi-step diffusion models have advanced both forward and inverse rendering, existing approaches often treat these problems independently, leading to cycle inconsistency and slow inference speed. In this work, we present Ouroboros, a framework composed of two single-step diffusion models that handle forward and inverse rendering with mutual reinforcement. Our approach extends intrinsic decomposition to both indoor and outdoor scenes and introduces a cycle consistency mechanism that ensures coherence between forward and inverse rendering outputs. Experimental results demonstrate state-of-the-art performance across diverse scenes while achieving substantially faster inference speed compared to other diffusion-based methods. We also demonstrate that Ouroboros can transfer to video decomposition in a training-free manner, reducing temporal inconsistency in video sequences while maintaining high-quality per-frame inverse rendering.

Framework Overview

The complete Ouroboros pipeline is organized around a cycle-consistent pair of single-step diffusion models, with an additional training-free extension for temporally consistent video inverse rendering.

Main Framework

Main Ouroboros framework overview
Main framework. Ouroboros couples inverse rendering and forward rendering through a cycle-consistent pair of single-step diffusion models.

Video Inference Framework

Video inference framework overview
Video inference framework. A training-free extension improves temporal consistency for video inverse rendering.

Results

We evaluate Ouroboros on both forward rendering and inverse rendering across diverse indoor and outdoor scenes. The visual comparisons below highlight two key observations: first, the cycle-consistent single-step design produces cleaner intrinsic estimates and more faithful reconstructions than RGB↔X in forward rendering; second, the inverse rendering model generalizes across scene types while preserving plausible albedo, normal, roughness, metallicity, and irradiance predictions.

Beyond per-image quality, Ouroboros also supports a training-free video extension that improves temporal consistency, which is reflected in the framework above and the stable appearance of the predicted intrinsic maps.

Forward Rendering

Given predicted intrinsic maps, Ouroboros synthesizes images that remain close to the original appearance while maintaining consistency with inverse rendering outputs. Compared with RGB↔X, our method yields more coherent material attributes and more faithful reconstructions.

Forward rendering example - Scene 1
Forward rendering example - Scene 2
Forward rendering example - Scene 3
Method RGB Albedo Normal Roughness Metallicity Irradiance Reconstruction
Ours
Ours RGB frame Ours Albedo Ours Normal Ours Roughness Ours Metallicity Ours Irradiance Ours Reconstruction
RGB↔X
RGB↔X RGB frame RGB↔X Albedo RGB↔X Normal RGB↔X Roughness RGB↔X Metallicity RGB↔X Irradiance RGB↔X Reconstruction

Inverse Rendering

The inverse rendering results show robust decomposition across both indoor and outdoor scenes. The predicted albedo, normal, roughness, metallicity, and irradiance maps remain structurally consistent across diverse inputs.

RGB Albedo Normal Roughness Metallicity Irradiance
Inverse rendering example test2 - RGB
Inverse rendering example test2 - Albedo
Inverse rendering example test2 - Normal
Inverse rendering example test2 - Roughness
Inverse rendering example test2 - Metallicity
Inverse rendering example test2 - Irradiance
Inverse rendering example test5 - RGB
Inverse rendering example test5 - Albedo
Inverse rendering example test5 - Normal
Inverse rendering example test5 - Roughness
Inverse rendering example test5 - Metallicity
Inverse rendering example test5 - Irradiance
Inverse rendering example test7 - RGB
Inverse rendering example test7 - Albedo
Inverse rendering example test7 - Normal
Inverse rendering example test7 - Roughness
Inverse rendering example test7 - Metallicity
Inverse rendering example test7 - Irradiance
Inverse rendering example MatrixCity test5 - RGB
Inverse rendering example MatrixCity test5 - Albedo
Inverse rendering example MatrixCity test5 - Normal
Inverse rendering example MatrixCity test5 - Roughness
Inverse rendering example MatrixCity test5 - Metallicity
Inverse rendering example MatrixCity test5 - Irradiance
Inverse rendering example MatrixCity test7 - RGB
Inverse rendering example MatrixCity test7 - Albedo
Inverse rendering example MatrixCity test7 - Normal
Inverse rendering example MatrixCity test7 - Roughness
Inverse rendering example MatrixCity test7 - Metallicity
Inverse rendering example MatrixCity test7 - Irradiance

BibTeX

@article{sun2025ouroboros,
  title={Ouroboros: Single-step Diffusion Models for Cycle-consistent Forward and Inverse Rendering},
  author={Sun, Shanlin and Wang, Yifan and Zhang, Hanwen and Xiong, Yifeng and Ren, Qin and Fang, Ruogu and Xie, Xiaohui and You, Chenyu},
  journal={arXiv preprint arXiv:2508.14461},
  year={2025}
}

Contact

For questions, feedback, or collaboration opportunities, please contact:
Email: shanlins@uci.edu, chenyu.you@stonybrook.edu