Albedo Estimation via Latent Bridge Matching
Carme Corbi, David Serrano-Lozano, Javier Vazquez-Corral, Maria Vanrell
cs.CV · cs.AI
Quick verdict
Efficient LBM albedo estimation with a physical consistency loss; well ablated, but headline gains are backbone-confounded and diffusion SOTA still leads.
Summary
The paper casts single-image albedo estimation as image-to-image translation with Latent Bridge Matching (LBM), which learns transport between the RGB-image and albedo distributions rather than from Gaussian noise. Because trajectories start at the observed image, the authors add a pixel-space reconstruction loss (I = A·S, optionally with a residual term) enforcing the intrinsic image formation model, and cap sampling at four steps (0.77 s/image versus 2.80 s for a 50-step diffusion baseline). Built on a frozen SDXL VAE and SDXL-initialized UNet and trained on InteriorVerse and Hypersim, LBM-AID beats identically-pipelined flow-matching and diffusion variants on most metrics (21.51 vs 19.92 PSNR on MIT). A conditioning study finds shading conditioning helps (IIW WHDR10 24.41→18.36) while surface normals hurt; the extended model also conditions the shading estimator on a first-pass albedo, giving the lowest reconstruction errors (MIT 0.0104, ARAP 0.0139, Hypersim 0.0250). Against SOTA on five benchmarks it achieves the best LMSE on MIT, but trails IntrinsicDiffusion on ARAP, DNF-Intrinsic on InteriorVerse (18.3 vs 21.9 PSNR), and RL-guided PRISM-X on IIW. Main limitations: the diffusion baselines use a weaker SD2.1 backbone, confounding the headline gains; no variance is reported; and the cost of the two-stage inference pipeline is never quantified.
Review complete
Assessment at a glance
5
risk flags identified
● high 0
● medium 3
● low 2
4 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- LBM is an effective generative framework for albedo estimation, cutting inference time to 0.77 s/img (about 72% below the 50-step SD baseline's 2.80 s) while improving accuracy over same-pipeline FM and diffusion variants.
- Because LBM trajectories are anchored to the input image, a pixel-space reconstruction loss can enforce the image formation model, improving physical consistency (e.g., Hypersim reconstruction MSE 0.0362→0.0305).
- Shading conditioning substantially improves albedo estimation (IIW WHDR10 24.41→18.36) whereas surface-normal conditioning consistently hurts, despite normals' geometric expressiveness.
- Conditioning the shading estimator on a first-pass predicted albedo (mutual conditioning) yields the best model, with the lowest reconstruction errors observed (MIT 0.0104, ARAP 0.0139, Hypersim 0.0250).
- The resulting model is competitive with multi-step diffusion SOTA — best LMSE on MIT (0.0030) — though it trails RL-guided variants on IIW and DNF-Intrinsic on InteriorVerse, a gap the authors attribute to synthetic-only training.
Strengths
Source-anchored transport enables pixel-level physical supervision
Because the LBM bridge (Eq. 4) starts at the observed image rather than Gaussian noise, the decoded albedo can be penalized with the formation-model loss L(I, A·S) (Eq. 8) — a constraint the Background section correctly argues noise-init diffusion/FM cannot apply. Table 5 shows Hypersim reconstruction MSE dropping from 0.0362 (without the loss) to 0.0305 (with it).
Controlled same-pipeline comparison isolates the formulation where backbone is fixed
Table 1: with identical SDXL backbones and training, LBM-AID beats FM-AID at equal 0.77 s/img cost — MIT PSNR 21.51 vs 19.92, IIW WHDR10 26.13 vs 32.54 — showing the bridge formulation itself adds accuracy at no extra inference cost.
Conditioning study yields a clear negative result plus a novel mutual-conditioning finding
Unusually honest SOTA positioning and limitations
Weaknesses
Headline gains confounded by the SD2.1→SDXL backbone swap
No variance, seeds, or significance testing; noisy ablations
Efficiency claim not carried through to the recommended model
Trails diffusion/RL SOTA on the largest benchmarks; some phrasing is generous
Cross-paper numbers and incomplete specification limit reproducibility
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
The paper (accepted at CIC 2026) targets single-image albedo estimation within Intrinsic Image Decomposition (IID). It argues that latent bridge matching (LBM) — which transports between the input-image distribution and the albedo distribution rather than from Gaussian noise — addresses three standing problems of generative IID: weak physical consistency, slow sampling, and generalization. Four contributions are claimed: LBM as an effective IID framework (up to 75% faster inference, 3-6 PSNR better than generative baselines); a pixel-space reconstruction loss enabled by source-anchored trajectories (reconstruction error reduced to 50%); a conditioning study showing shading beats surface normals; and a two-stage pipeline whose best variant also conditions the shading estimator on a first-pass albedo. An extended section benchmarks the best model on five datasets (MIT, IIW, ARAP, InteriorVerse, Hypersim).
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 4 strengths & 5 weaknesses, analyzed in detail
- 5 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is Latent Bridge Matching and why is it used for albedo estimation?
Latent Bridge Matching (LBM) is a generative image-to-image framework that learns transport between two image distributions instead of from Gaussian noise. For albedo estimation the transport goes from RGB images to albedo maps, so predictions stay anchored to the input pixels, sampling needs only a few steps, and a pixel-level physical consistency loss (image = albedo × shading) can be applied — things standard diffusion or flow matching cannot easily do.
How fast is the proposed model compared to diffusion-based intrinsic decomposition?
The single LBM albedo model runs at about 0.77 seconds per image, roughly 72% faster than the paper's 50-step Stable Diffusion baseline (2.80 s). The recommended final version is a two-stage pipeline (a shading estimator plus the albedo model, with a first pass), whose total runtime the paper does not report.
Does this method beat diffusion models like Marigold or IntrinsicDiffusion?
It depends on the benchmark. It achieves the best LMSE on MIT and beats earlier deep-learning baselines on ARAP, but trails IntrinsicDiffusion on ARAP, DNF-Intrinsic on InteriorVerse (18.3 vs 21.9 PSNR), Marigold and PRISM-X on Hypersim, and the RL-guided PRISM-X on IIW. Its main advantages are inference speed and physical consistency of the decomposition.
Is code available for this paper?
Yes — the paper links a GitHub repository (CVC-Color/albedoLBM). Training relies on the public synthetic InteriorVerse and Hypersim datasets; evaluation uses MIT, IIW, ARAP, and held-out splits of the two synthetic sets.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Figure contents (Figs. 1-7) are not present in the text extraction; qualitative judgments in this review rest on captions and the authors' own descriptions only.
- In-text cross-references appear as empty 'Section .' throughout the extraction; whether this reflects unresolved references in the PDF or an extraction artifact cannot be determined from the text.
- Mathematical notation is partially garbled by PDF extraction (e.g., Eqs. 4 and 7); formulas were interpreted from surrounding context.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.