Geng Integrity Report: Structure-Guided Transformer with 3D Gaussian Splatting for Deblurred Single-View 3D Reconstruction

Academic fraud report · Geng Detector

Summary

Verdict: Highly suspicious. This report flags a Computer Science and Application paper (DOI 10.12677/csa.2026.161016) titled '基于结构引导 Transformer 的单视图三维重建去模糊方法' by Zhang Yuanmeng, Lin Lixia, and Cao Peng (Beijing Institute of Graphic Communication). Several anomalies undermine confidence in the reported experiments. Most critically, Table 1 (Restormer + SI baseline) and Table 2 (ablation variant 'Ours w/o SGFN, L_grad + SI') report identical PSNR/SSIM/LPIPS values to two decimal places across architectures that differ only in label, an outcome essentially impossible from independent training runs. Second, all metrics are reported as point estimates without standard deviations, inconsistent with standard practice on GoPro-scale benchmarks. Third, the gradient-consistency loss in Eq. (1) is written as a signed sum of gradient differences, allowing positive and negative errors to cancel, rather than a non-negative L1/L2 norm, raising serious concerns about mathematical validity and code-to-paper consistency. Additional soft concerns include a suspiciously rapid submission-to-acceptance timeline (7 Dec 2025 to 9 Jan 2026) and the journal's established reputation. Figures could not be analyzed for pixel-level duplication. Confidence in findings 1–3 is high; figures and code remain unverified.

Verdict

🟠 Highly suspicious. Multiple statistical and mathematical inconsistencies suggest that reported results may have been fabricated, copied, or back-filled rather than produced by genuine training runs. Independent verification via code and training logs is strongly recommended.

Key findings

Identical values across architecturally distinct models (Finding 1, severity 🔴): In Table 1, the 'Restormer + SI' baseline reports PSNR 17.46, SSIM 0.84, LPIPS 0.156. In Table 2, 'Ours w/o SGFN, L_grad + SI' reports the exact same PSNR 17.46, SSIM 0.84, LPIPS 0.156. Two differently structured deep models producing numerically identical metrics to two decimal places is effectively impossible from independent experiments.
Missing variability statistics (Finding 2, severity 🟠): All PSNR/SSIM/LPIPS results in Tables 1 and 2 are reported as single point estimates with no standard deviation or standard error, atypical for image-restoration benchmarks such as GoPro.
Mathematically invalid loss function (Finding 3, severity 🔴): Equation (1) defines L_grad as a signed sum of gradient differences, allowing positive and negative terms to cancel. A valid loss must be non-negative; the standard form uses L1 or L2 norm, e.g., ||∇Î − ∇I_gt||_1. This indicates either a fundamental mathematical error or a mismatch between paper and code.
Figure-level checks not performed (Finding 4, severity 🟡): Figures 1 and 2 could not be analyzed for splicing or image reuse without image files.
Rapid publication timeline (Finding 5, severity 🟠): Received 2025-12-07, accepted 2026-01-09; combined with the journal's (Hans Publishers / Computer Science and Application) reputation, this raises concerns about peer-review rigor.

Evidence highlights

DOI: 10.12677/csa.2026.161016
Journal: Computer Science and Application / 计算机科学与应用 (Hans Publishers)
Title: 基于结构引导 Transformer 的单视图三维重建去模糊方法
Authors / Affiliation: 张媛梦，林立霞，曹鹏 — 北京印刷学院信息工程学院
Year: 2026 (published online January 2026)
Submission-to-acceptance span: ~33 days (2025-12-07 → 2026-01-09)
Numeric match (Tables 1 vs 2): Restormer + SI == Ours w/o SGFN, L_grad + SI → PSNR 17.46, SSIM 0.84, LPIPS 0.156
Loss formula anomaly: L_grad = ∇_x Î − ∇_x I_gt + ∇_y Î − ∇_y I_gt (signed, cancellation-prone)
Metrics format: Point estimates only; no ± SD / SE reported

Notes

The allegation of image reuse in Figures 1–2 is not investigated here; no image access was available. Pixel-level forensics remain pending.
All quantitative anomalies (Findings 1–2) are fully reproducible from the text and do not require image analysis.
The mathematical issue in Finding 3 is presented as written; alternative interpretations (e.g., an implicit norm, a typesetting omission) cannot be ruled out solely from the PDF text.
Final determination of misconduct requires an institutional or editorial investigation, including inspection of training logs, code repositories, and raw experimental records. PubPeer commentary and direct author contact are appropriate next steps.
This report is AI-assisted and intended for academic discussion only.