PixlRelight: Physically Controllable Single-Image Relighting

Model Overview

PIXLRelight is a feed-forward, transformer-based neural rendering framework for physically controllable single-image relighting. Developed by Miguel Farinha and Ronald Clark at the University of Oxford, it bridges physically based rendering (PBR) and learned image synthesis via shared intrinsic conditioning.

By using per-pixel affine modulation, PIXLRelight preserves fine-grained source details, textures, and scene geometry while achieving ultra-fast feed-forward inference (<0.1s per image).

---

Key Features

• Physically Controllable Relighting: Enables exact lighting changes on a single image using target HDRI maps or synthetic PBR lighting controls.
• Dual Conditioning Modes: Accepts either real target reference photographs (via Marigold-IID-Lighting) or path-traced Blender Cycles renders.
• Ultra-Fast Feed-Forward Inference: Performs single-image relighting in under 0.1 seconds per image on modern GPUs.
• Detail-Preserving Affine Modulation: Employs per-pixel affine transformations to preserve original textures and geometric boundaries.
• Integrated Intrinsic Estimation: Incorporates Depth Anything 3 for geometry recovery and Marigold-IID-Appearance for material estimation.

---

Verified Project Links

• Project Website: https://mlfarinha.github.io/pixl-relight/
• arXiv Paper: https://arxiv.org/abs/2605.18735
• GitHub Repository: https://github.com/mlfarinha/pixlrelight
• Hugging Face Model: https://huggingface.co/mlfarinha/pixlrelight

---

Performance & Benchmarks

• Relighting Quality (PSNR/SSIM): Outperforms prior baseline methods (DiffusionRenderer, UniRelight, RGBX) on synthetic and real benchmarks.
• Perceptual Quality (LPIPS): Delivers state-of-the-art perceptual scores at sub-100ms inference speeds.