- Rating
- 3(+1)
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
The paper addresses cover-guided Live Photo super-resolution: restoring the entire low-quality Live Photo video using its high-quality native cover, whereas prior Live Photo methods (LiveMoments, Live2K) output a single image.
LiveVSR adapts Wan2.1-1.3B into a one-step diffusion restorer with two components:
- In-context framework: jointly processes cover and video tokens with bidirectional self-attention, plus flow-guided local retrieval (RAFT-based, confidence-weighted mixing of local and global cover attention).
- Correspondence-guided codebook injection: uses hard top-1 matching (trained with STE) to inject discrete features from a codebook pretrained on HQ images, intended to mitigate residual cover degradation.
Training mixes clean and mildly degraded references. Experiments on UDM10, SPMCS, YouHQ40 (reference = mildly degraded GT middle frame) and a new 40-clip RealLive set report large fidelity/perceptual gains on synthetic data.
Critical strengths
- Practical and well-motivated task. Extending cover guidance from a single frame to the full Live Photo video is a natural, useful step, and the planned RealLive release could benefit the community.
- Strong fidelity gains when a good reference is available. On synthetic benchmarks, LiveVSR substantially outperforms all baselines in PSNR/SSIM/LPIPS/DISTS (e.g., UDM10 PSNR 29.18 vs. 26.62; LPIPS 0.096 vs. 0.258), and the ablations (Tab. 2b) show that joint self-attention uses the reference much better than cross-attention or ReferenceNet, with a modest one-step 1.3B model.
Critical weaknesses
1. Key claims are not supported by the results.
- Temporal consistency. The contribution list (Sec. 1) claims that LiveVSR outperforms existing methods in temporal consistency, but in Table 1 its ranks only 3rd on all synthetic sets (e.g., UDM10 1.105 vs. 0.781 for SparkVSR*) and 7th of 10 on RealLive (3.990 vs. 2.577 for ERVSR). The promised temporal analysis "in Appendices C and D" does not exist (C is efficiency, D is visual results). The Table 1 caption ("outperforms ... across various datasets and metrics") is also overstated (e.g., DOVER ranks 2nd–4th; several relative NR metrics on YouHQ40/RealLive favor LiveMoments).
- Codebook injection. Presented as a core contribution for handling degraded covers, but its gain is marginal (+0.235 dB PSNR over +Flow, with slightly worse DISTS and DOVER in Tab. 2a; all codebook variants worsen CLIPre in Tab. 3b). There is no experiment varying cover degradation and no visualization (e.g., retrieved codes/correspondences) showing that it actually suppresses cover artifacts.
2. Evaluation fairness and numerical inconsistencies.
- SparkVSR is not given the same cover. SparkVSR, itself a keyframe-reference method, uses a PiSA-SR-generated reference instead. Since LiveVSR's NoRef variant (25.59 dB) is below DOVE (26.53 dB), its lead comes almost entirely from the reference. Moreover, on synthetic sets that reference is the GT middle frame degraded with exactly the training degradation (strength 0.10), i.e., a near-oracle, train-test-matched reference. Without giving SparkVSR the same cover, the comparison does not isolate the contribution of the proposed method.
- Table inconsistencies. Attention and Ours have identical CLIPre (0.1997) in Tab. 2b despite very different other metrics; the full model's DOVER is 0.7986 in Tab. 1 but 0.8014 in Tab. 2a.
Minor (presentation/reproducibility)
- Broken references ("Appendix ??", "detailed in .", "Figs. 5, ??, and 6").
- Duplicated sentences in Sec. 4.3.
- Undefined footnotes in Tab. 5, whose runtime is from an "earlier pipeline".
- Missing hyperparameters (, , , neighborhood size, codebook size, stage-2 LR, iterations).
Questions for authors
- Please reconcile the temporal-consistency claim with Table 1 (especially on RealLive), and provide the temporal analysis referenced in Appendices C/D.
- Please correct/explain the identical CLIPre values in Tab. 2b and the DOVER mismatch between Tab. 1 and Tab. 2a.
- Fairness: results for SparkVSR using the same cover as its keyframe reference, and PSNR as a function of frame distance from the reference.
- Codebook: results with and without codebook injection under varying cover degradation (clean, 0.10, stronger, real-cover-like), plus visualizations showing that it suppresses cover artifacts.