OpenNovelty.org

PI-Light: Physics-Inspired Diffusion for Full-Image Relighting

ICLR 2026 Conference SubmissionAnonymous Authors
OpenReview Score: 6.7Download Report PDF
Diffusion modelRelightingInverse renderingNeural forward rendering
Abstract:

Full-image relighting remains a challenging problem due to the difficulty of collecting large-scale structured paired data, the difficulty of maintaining physical plausibility, and the limited generalizability imposed by data-driven priors. Existing attempts to bridge the synthetic-to-real gap for full-scene relighting remain suboptimal. To tackle these challenges, we introduce Physics-Inspired diffusion for full-image reLight (π\pi-Light, or PI-Light), a two-stage framework that leverages physics-inspired diffusion models. Our design incorporates (i) batch-aware attention, which improves the consistency of intrinsic predictions across a collection of images, (ii) a physics-guided neural rendering module that enforces physically plausible light transport, (iii) physics-inspired losses that regularize training dynamics toward a physically meaningful landscape, thereby enhancing generalizability to real-world image editing, and (iv) a carefully curated dataset of diverse objects and scenes captured under controlled lighting conditions. Together, these components enable efficient finetuning of pretrained diffusion models while also providing a solid benchmark for downstream evaluation. Experiments demonstrate that π\pi-Light synthesizes specular highlights and diffuse reflections across a wide variety of materials, achieving superior generalization to real-world scenes compared with prior approaches.

Disclaimer
This report is AI-GENERATED using Large Language Models and WisPaper (A scholar search engine). It analyzes academic papers' tasks and contributions against retrieved prior work. While this system identifies POTENTIAL overlaps and novel directions, ITS COVERAGE IS NOT EXHAUSTIVE AND JUDGMENTS ARE APPROXIMATE. These results are intended to assist human reviewers and SHOULD NOT be relied upon as a definitive verdict on novelty.
NOTE that some papers exist in multiple, slightly different versions (e.g., with different titles or URLs). The system may retrieve several versions of the same underlying work. The current automated pipeline does not reliably align or distinguish these cases, so human reviewers will need to disambiguate them manually.
If you have any questions, please contact: mingzhang23@m.fudan.edu.cn

Overview

Taxonomy

Core-task Taxonomy Papers
50
3
Claimed Contributions
30
Contribution Candidate Papers Compared
1
Refutable Paper

Research Landscape Overview

Core task: full-image relighting using physics-inspired diffusion models. The field has organized itself around several complementary strategies for manipulating illumination in images. At the highest level, one branch pursues intrinsic decomposition and physics-based rendering integration, explicitly factoring scenes into reflectance and shading components before applying physically grounded transformations. A second branch adopts direct diffusion-based relighting, leveraging generative priors to synthesize new lighting conditions end-to-end without explicit decomposition. Temporal consistency for video relighting addresses the challenge of maintaining coherent illumination across frames, while low-light image enhancement via diffusion models focuses on recovering detail and color in underexposed captures. Finally, domain-specific applications tackle specialized scenarios such as portrait relighting, underwater imaging, and multi-view synthesis, each imposing unique constraints on how lighting can be modeled and controlled. Within the direct diffusion-based relighting branch, a particularly active line of work targets full-scene and multi-view relighting, where methods must handle complex spatial layouts and viewpoint variations. PI-Light[0] exemplifies this direction by incorporating physics-inspired constraints directly into the diffusion process, aiming to balance generative flexibility with physical plausibility. Nearby efforts such as Lightlab[1] and Multi-Illumination Synthesis[17] explore similar full-scene settings but differ in their treatment of multi-illumination priors and the degree of user control over lighting parameters. LumiNet[31] and Comprehensive Relighting[46] further illustrate the trade-offs between end-to-end learning and modular design: some approaches prioritize seamless integration of lighting cues into a single network, while others decompose the problem into sub-tasks that can be individually optimized. Across these works, open questions remain about how to enforce consistency across viewpoints, how much physical modeling to embed versus learn, and how to scale these techniques to diverse real-world scenes.

Claimed Contributions

PI-Light: Physics-Inspired Diffusion Framework for Full-Image Relighting

The authors propose a two-stage diffusion-based framework for full-image relighting that incorporates batch-aware attention for consistent intrinsic predictions, a physics-guided neural rendering module enforcing physically plausible light transport, and physics-inspired losses that regularize training toward physically meaningful solutions to enhance generalization to real-world scenes.

10 retrieved papers
Physics-Inspired Light Transport Prior for Neural Forward Rendering

The authors introduce physics-inspired losses (diffuse shading loss and physical-based shading loss) that regularize the neural forward rendering module to follow physically plausible light transport principles, enabling the model to learn correct light transport with less data and computation while improving generalization.

10 retrieved papers
Curated Dataset of Objects and Scenes Under Controlled Lighting

The authors construct a new dataset featuring diverse objects from Objaverse and curated scenes from BlenderKit, all rendered under controlled lighting conditions with ground-truth intrinsic properties, addressing data scarcity in full-image relighting research and enabling comprehensive downstream benchmarking.

10 retrieved papers
Can Refute

Core Task Comparisons

Comparisons with papers in the same taxonomy category

Contribution Analysis

Detailed comparisons for each claimed contribution

Contribution

PI-Light: Physics-Inspired Diffusion Framework for Full-Image Relighting

The authors propose a two-stage diffusion-based framework for full-image relighting that incorporates batch-aware attention for consistent intrinsic predictions, a physics-guided neural rendering module enforcing physically plausible light transport, and physics-inspired losses that regularize training toward physically meaningful solutions to enhance generalization to real-world scenes.

Contribution

Physics-Inspired Light Transport Prior for Neural Forward Rendering

The authors introduce physics-inspired losses (diffuse shading loss and physical-based shading loss) that regularize the neural forward rendering module to follow physically plausible light transport principles, enabling the model to learn correct light transport with less data and computation while improving generalization.

Contribution

Curated Dataset of Objects and Scenes Under Controlled Lighting

The authors construct a new dataset featuring diverse objects from Objaverse and curated scenes from BlenderKit, all rendered under controlled lighting conditions with ground-truth intrinsic properties, addressing data scarcity in full-image relighting research and enabling comprehensive downstream benchmarking.

Table of Contents

PI-Light: Physics-Inspired Diffusion for Full-Image Relighting | Novelty Validation