Geometry-Driven Shadow Harmonisation for Composited Faces: A Multiplicative, Albedo-Preserving Relighting Pipeline
cs.CV
Quick verdict
Provably bounded, chromaticity-preserving shadow-injection operator for face composites; transparent and useful, but evaluated only on a synthetic heightfield.
Summary
The paper addresses a specific failure of face compositing: a donor face captured under flat, frontal studio light transplanted onto a body lit by directional scene light. Instead of neural relighting or colour transfer, it injects form shadow through a multiplicative gain field g in [0.82, 1], applied channel-uniformly in linear RGB, so pixels can only darken and chromaticity is provably preserved (Proposition 1). Geometry comes from 468 MediaPipe landmarks rasterised into a depth buffer, from which normals, a cavity term, and screen-space ray-marched cast shadows are derived. Key-light direction is estimated mainly from host-side cues (body, background, hair halo), with on-face cues downweighted because they recover the donor's lighting. Shadow magnitude is set by a three-parameter transfer (τ=0.90, σ=0.45, gmin=0.82) with 75th-percentile normalisation, and is deliberately not matched to the host. On a synthetic analytic face heightfield with a known light direction, the default configuration modifies 56.5% of face pixels (mean gain 0.938; 0.890 on modified pixels; 3.4% at the floor). The authors state the limitations frankly: no experiments on real composites, no baselines, no perceptual evaluation, an unvalidated cue-fusion estimator, and double-shadowing of donors that already carry shading.
Review complete
Assessment at a glance
4
risk flags identified
● high 0
● medium 2
● low 2
5 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
- A channel-uniform multiplicative gain bounded in [gmin, 1] provably cannot brighten pixels, shift linear chromaticity, or exceed a (1−gmin) radiance reduction (Proposition 1).
- On swapped faces, key-light estimation should be dominated by host-side cues because on-face cues recover the donor's studio lighting.
- At defaults (τ=0.90, σ=0.45, gmin=0.82) the operator modifies 56.5% of face pixels on the analytic heightfield with mean gain 0.938, engaging the floor on only 3.4% of pixels.
- The bounded operator enables unattended operation because worst-case damage from a wrong mesh or light direction is a misplaced shadow of at most 1−gmin, not identity alteration.
Strengths
Formal operator guarantees
Proposition 1 gives closed-form invariants — no brightening, exact linear-RGB channel-ratio preservation, deviation bounded by 1−gmin — and Remark 1 correctly states when the bound is voided (smoothing without re-clip, non-zero tint).
Unusually honest scoping
Sec. 5 explicitly disclaims ground truth, real composites and baselines; Sec. 6 contains dedicated 'What the constraints cost' and 'What the evaluation does not show' paragraphs; the abstract's numbers match Table 2 exactly.
Complete implementation disclosure
Sensible in-scope ablations
Correct diagnosis of donor-light circularity
Weaknesses
No evaluation in the target domain
Key-light estimator unvalidated
No baselines whatsoever
Magnitude decoupled from the host
Proxy construction under-specified
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
The paper targets a narrow, well-defined failure mode: donor faces photographed under flat frontal studio light composited onto directionally lit bodies. The claimed contribution is not better relighting but a *bounded* one — a per-pixel multiplicative gain g∈[gmin,1] with three provable invariants (Proposition 1: no brightening; exact linear-RGB channel-ratio invariance; deviation bounded by 1−gmin=0.18). Four contributions are listed: landmark rasterisation at 512², host-side cue-weighted key-light estimation, a shadow field combining wrapped Lambertian shading, screen-space cast shadows and cavity AO with 75th-percentile normalisation, and a three-parameter gain transfer. The abstract's quantitative claims (56.5% modified pixels, mean gain 0.938, 3.4% at floor) match Table 2 exactly. Crucially, Sec. 5 states: 'No ground-truth relit pair, no real composite, and no baseline comparison is claimed' — the paper never overreaches its evidence.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 4 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
Does this method change the skin tone or identity of the face?
No. By construction it only multiplies linear RGB by a channel-uniform gain between 0.82 and 1.0, so chromaticity is mathematically preserved and darkening is capped at 18%; no pixels are synthesised or repainted.
Was the method tested on real face swaps?
No. All published numbers come from a synthetic analytic face heightfield with a known light direction. Experiments on real composites, user studies and baseline comparisons are explicitly left as future work by the authors.
Is code available for this paper?
Yes. The paper links an implementation at github.com/vijeshkpaei/geometry-driven-shadow-harmonisation, and Appendix A documents every default hyperparameter.
What are the default parameters of the pipeline?
τ=0.90 (lit gate), σ=0.45 (strength), gmin=0.82 (gain floor), depth scale k=1.15, 512×512 working crop, Lambert wrap w=0.28 with exponent γ=1.4, and 48 ray-march steps over at most 0.22 crop widths.
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.
- Figures were available only through captions; the curves in Figs. 4-6 could not be visually inspected.
- Preprint posted to arXiv (Sep 2026); no peer-review status is indicated.
- The linked GitHub repository was not inspected; code availability is as stated in the paper.
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.