BRDFusion¶

Development line: project:brdfusion · thread brdfusion
Last event: 2026-06-22 · 1 dated since 2026-06-22 · Researched: 2026-09-05 · confidence: medium

What it is¶

BRDFusion is an open research pipeline for turning urban video into a controllable 3D Gaussian/PBR scene, closer to UrbanIR plus diffusion refinement than to a general video generator.

• Geometry, materials, and HDR lighting: reconstructs them from video.
• Novel views and relighting: renders both through PBR.
• Rendered video: refines frames with DiffusionRenderer.
• Scene edits: supports night simulation, local lights, headlights, and dynamic-object edits.

Tested on Ubuntu 22.04 with an RTX A6000; Gen. Render can exceed 24 GB of VRAM. We use it for reproducible research and driving-scene simulation on a Linux GPU stack, not as a lightweight single-model workflow.

Development line¶

• 2026-06-22 — BRDFusion project resources were published. On 2026-06-22, BRDFusion linked its project page, source repository, and checkpoint collection. This release makes the implementation and model artifacts public so researchers can locate and use them.

What changed¶

• 2026-06-15 — arXiv v1 introduced BRDFusion’s hybrid physics-and-generation method for urban-scene inverse and forward rendering.
• 2026-06-22 — The project page, public code, and pretrained-checkpoint resources were linked as a usable research release.

How to use this¶

From 2026-06-22, we can evaluate or reproduce the project starting from the linked BRDFusion project page, source repository, and checkpoint collection.

1. Check the documented Linux, compiler, CUDA, and GPU prerequisites before installing; the project uses separate brdfusion and cosmos-predict1 environments. — https://github.com/shigon255/BRDFusion
2. Clone recursively, build both Conda environments, authenticate with Hugging Face for DiffusionRenderer weights, and supply the required SMPL asset. — https://github.com/shigon255/BRDFusion
3. Download a supplied preprocessed scene and matching one-camera checkpoint, unpack it under ckpt/, then run bash scripts/stage_ckpt.sh. — https://github.com/shigon255/BRDFusion
4. Run tools/run_pipeline.py with --stage render for PBR output; use the matching scene, camera, frame range, and render target for optional --stage gen_render. — https://github.com/shigon255/BRDFusion
5. For relighting or scene edits, set the documented environment variables or JSON specs and invoke scripts/applications/render.sh against a staged checkpoint. — https://github.com/shigon255/BRDFusion

Best practices¶

• Start with the supplied data and pretrained checkpoints for inference or evaluation before attempting training or new Waymo-scene preprocessing. — https://github.com/shigon255/BRDFusion
• Download only the scene subset needed for an experiment; training, rendering, and metrics operate per scene. — https://github.com/shigon255/BRDFusion
• Keep dataset, scene or path, camera setting, frame range, and render target identical between render and Gen. Render. — https://github.com/shigon255/BRDFusion
• Use --dry_run to inspect selected commands and the pipeline manifest before expensive work; existing stage outputs are skipped unless --rerun_existing is supplied. — https://github.com/shigon255/BRDFusion
• Use NO_PBR=1 and PRINT_CMD=1 when checking geometry, camera paths, or generated commands before a full application render. — https://github.com/shigon255/BRDFusion

Superseded by this¶

• Nothing marked obsolete yet.

Still unknown¶

• We found no dated primary release note for a code or checkpoint change on 2026-06-22 itself; the date establishes the link event, but not a separate upstream version.
• The repository documents 35 commits without a dated commit history in available research access, so we cannot assign current README capabilities to a post-22-June development event.
• The source code is MIT-licensed, while the checkpoint card is labelled License: other; checkpoint and bundled third-party use terms need review before commercial deployment.
• We found no independent reproduction, runtime-performance measurement, or production deployment evidence.

Sources¶

Agent brief¶

• Subject: project:brdfusion, thread brdfusion, 1 dated events 2026-06-22 → 2026-06-22.
• Practical note: From 2026-06-22, use the linked BRDFusion project page, source repository, and checkpoint collection to evaluate or reproduce the project.
• Confidence: medium. Dated supersedes above are the authority for what is obsolete.