{"id":"61a300ce-8ae2-4838-9b5d-6093930317a8","arxiv_id":"2506.13355","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"DicFace replaces discrete codebook selection in face restoration with Dirichlet-distributed convex combinations and a spatio-temporal Transformer, yielding state-of-the-art temporal coherence on VFHQ.","lead":"This paper introduces DicFace, a video face restoration model that treats codebook entries as continuous Dirichlet-distributed mixtures instead of discrete codes, and applies a spatio-temporal Transformer across frames. It reports state-of-the-art results on the VFHQ benchmark for blind restoration, inpainting, and colorization, with notable gains in temporal stability.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The paper's own ablation contradicts its headline temporal-coherence result: Table 1 reports Ours TLME=1.091, while Table 3 reports TLME=1.107 for the same α=1.0 configuration and TLME=1.091 without the KL term.","rationale":"The most load-bearing condition for the stated contribution is not merely high PSNR but that the Dirichlet/variational mechanism is what produces temporal coherence. The authors present Table 3 as evidence, yet the table contradicts the headline: the α=1.0 row has the same PSNR, AKD, and FVD as Table 1's Ours row but a different TLME, and the no-KL row has the same TLME as Table 1's Ours. I therefore focus on this internal numeric inconsistency rather than on the reader's reparameterization concern, because the latter can be settled by inspecting the released training code (for example, checking whether torch.distributions.Dirichlet.rsample is used and whether gradients flow to the concentration parameters), whereas the table inconsistency is not resolvable from the text and directly affects the central temporal-coherence claim. The finding is not an accusation of misconduct; it is a concrete reproducibility check that must pass before the reported temporal improvement can be credited to the method. This supports the reader's CONDITIONAL verdict and does not move it to acceptance or rejection.","tokens_in":11547,"tokens_out":10723,"duration_ms":112357,"concrete_test":"Run the released checkpoint on VFHQ-Test with α=1.0 and with the KL loss removed, using the paper's stated degradation pipeline and metric code, and compare PSNR and TLME against Tables 1 and 3. If α=1.0 yields TLME=1.107 while Table 1 gives 1.091, the headline temporal result is internally inconsistent; if the w/o-KL configuration has TLME equal to or better than α=1.0, the central attribution of temporal coherence to the Dirichlet prior fails.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The paper's central claim is temporally coherent restoration, supported by the headline TLME of 1.091 in Table 1. In the Dirichlet-prior ablation (Table 3), the configuration corresponding to Ours (α=1.0; identical PSNR 29.099, AKD 2.093, FVD 336.015 as Table 1) is reported with TLME=1.107, while the row with the KL term removed ('w/o') reports TLME=1.091. Thus either Table 1 or Table 3 is incorrect, and as printed the data show that the Dirichlet KL term does not improve temporal stability. The text's claim that ablations validate the importance of Dirichlet-based variational modeling for temporal coherence is therefore unsupported by the reported numbers. A related arithmetic error appears in Table 4's '5.7% reduction in AKD versus the Average baseline': the actual relative reduction is (2.128−2.093)/2.128 ≈ 1.6%, reinforcing that the numeric claims need independent recomputation before the SOTA temporal improvement can be accepted.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes DicFace, a video face restoration method that relaxes the discrete codebook assignments of a pretrained VQ-VAE (CodeFormer) into continuous Dirichlet-distributed convex combinations, modeled by a spatio-temporal Transformer, and trained with an ELBO plus LPIPS loss. Claims state-of-the-art blind video face restoration, inpainting, and colorization on VFHQ-Test, with improved temporal coherence as measured by TLME, FVD, and landmark stability. The manuscript includes quantitative tables, qualitative figures, and a series of ablations on the Dirichlet prior and code aggregation strategy, and it reports that source code is publicly available.","tokens_in":11833,"tokens_out":3892,"duration_ms":39354,"significance":"If the claims hold after the inconsistencies and missing specifications are resolved, the core idea is valuable: it offers a principled way to reuse strong image-domain codebook priors for video while explicitly modeling temporal smoothness in latent code selection. The empirical story is appealingly consistent across three tasks, and the decision to build on an external, cited pretrained codebook rather than fitting a bespoke temporal metric is a reasonable design. The availability of open-source code is also a strength. However, significance is currently conditional on correcting the Dirichlet KL formula, specifying the differentiable sampling procedure, and reconciling the conflicting TLME numbers in Tables 1 and 3.","major_comments":[{"comment":"The training objective depends on 'differentiably sampling w_l from q_theta(w|x)', but the manuscript never specifies how gradients are obtained through Dirichlet sampling. The standard location-scale reparameterization does not apply to the Dirichlet distribution, and the chosen estimator (e.g., implicit reparameterization, Gamma-based reparameterization with a straight-through variant, or the softmax-Gaussian trick) materially affects the ELBO gradient and training behavior. This is a load-bearing implementation detail that must be stated. In addition, Eq. (10) writes log p_theta(y|x,w_l) = |y - y_pred_l| + C, omitting the required negative sign for a Laplacian log-likelihood; as printed, the 'reconstruction' term is positive and maximization of Eq. (7) would be ill-posed. Please correct the sign and specify the sampling/reparameterization scheme.","section":"Section 3.3, Eq. (10)"},{"comment":"The closed-form KL divergence appears to have a typo: the final digamma term uses psi(sum_k alpha_{i,j,k}) but the standard KL between two Dirichlets uses psi(sum_k alpha_hat_{i,j,k}). The difference is not a constant with respect to the predicted alpha_hat, because it is multiplied by (alpha_hat_{i,j,k} - alpha_{i,j,k}). If implemented as written, this would change the gradient of the KL term. Please verify the formula and correct it, or clarify the convention if a different parametrization is intended.","section":"Section 3.2, Eq. (8)"},{"comment":"There is a direct numerical contradiction in the central temporal-coherence metric. Table 1 reports 'Ours' with TLME = 1.091, PSNR = 29.099, AKD = 2.093, and FVD = 336.015. Table 3 reports the identical configuration (alpha = 1.0) with the same PSNR, AKD, and FVD values but TLME = 1.107, while the row with the KL term removed ('w/o') reports TLME = 1.091. Thus, as printed, the ablation shows that removing the Dirichlet KL term leaves TLME unchanged from the headline value and that the alpha = 1.0 model is worse on TLME. This directly undermines the paper's claim that the Dirichlet variational modeling improves temporal coherence. The numbers must be reconciled, and the temporal-coherence conclusion should be based on the corrected table.","section":"Tables 1 and 3"},{"comment":"The manuscript repeatedly defers essential experimental details to an Appendix that is not present in the submitted version: the degradation pipeline for VFHQ, the definitions of all metrics (especially TLME, AKD, and FVD), the evaluation protocols, out-of-domain results, and additional ablations (temporal Transformer, codebook size, number of Monte Carlo samples) are all listed as 'in Appendix'. Without these definitions and protocols, the quantitative claims in Tables 1 and 2 cannot be reproduced or fully assessed. Please include the missing appendix or move the definitions into the main text.","section":"Section 4.1 and Appendix references"},{"comment":"The text states that the proposed method gives a '5.7% reduction in AKD versus the Average baseline', but the table values give (2.128 - 2.093)/2.128 ≈ 1.6%. The same paragraph's '4.1%' FVD reduction is also not directly reproducible from the table: (342.141 - 336.015)/342.141 ≈ 1.8%, not 4.1%. Please recompute all relative improvements and ensure that every stated percentage matches the reported table values.","section":"Section 4.3, Table 4 and surrounding text"}],"minor_comments":[{"comment":"Reference [22] is cited as CodeFormer in several places, but the listed reference is 'Codeformer: A GNN-Nested Transformer Model for Binary Code Similarity Detection', which is a different paper on binary code analysis. The correct CodeFormer face-restoration reference (Zhou et al., NeurIPS 2022) should be cited.","section":"References"},{"comment":"The sentence 'Our framework produces correct eyes and mouses' contains a typo; 'mouses' should be 'mouths'.","section":"Section 4.2, Inpainting paragraph"},{"comment":"The notation '2H alternating attention blocks' is ambiguous because H is already used for image height. Please use a different symbol, such as N_blocks, for the number of alternating block pairs.","section":"Section 3.4"},{"comment":"The inference description says each sliding window consists of 5 consecutive frames, but the method uses R frames in training; please clarify whether R is fixed to 5 and whether the temporal attention window at inference matches the training window size.","section":"Section 3.5, Inference"},{"comment":"The variance visualization in Figure 8(b) is informal; if this is used to support the claim of stable weights over pixels, please provide a quantitative summary (e.g., mean and standard deviation over pixels and frames) rather than only a shaded plot.","section":"Figure 8"},{"comment":"Equation (9) sets lambda_1 = -1.0 and lambda_2 = 1.0, but the text says this 'prioritizes reconstruction fidelity'; please clarify the sign convention explicitly, since Eq. (7) already defines the ELBO as a quantity to be maximized while L_total is minimized.","section":"Section 3.3, Loss definition"}],"recommendation":"major_revision","confidential_remarks":"The manuscript appears technically promising but is not yet complete. The missing appendix is a substantive issue for a submission that bases its claims on three tasks and multiple metrics. The numerical inconsistency between Tables 1 and 3 on TLME is the most serious concern, since temporal coherence is the paper's primary contribution. The incorrect citation for CodeFormer also suggests that reference checking was rushed. I would encourage the editor to ask for a revised version with the appendix included, the equations corrected, the tables reconciled, and the sampling procedure specified before sending it back to review."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"First thing you should know: the paper has a genuinely new idea—relaxing VQ codebook assignments to Dirichlet distributions for video face restoration—but the headline temporal-coherence result unravels when you cross-check the ablation tables. The method is worth a serious look, but the manuscript needs major cleanup before the numbers can be trusted.\n\nWhat's good: the core approach is sensible. Instead of hard codebook lookup per frame, they predict a Dirichlet posterior over codebook weights, sample from it, and decode convex combinations. The spatio-temporal Transformer is an obvious but effective way to share context across frames. On VFHQ they report solid gains over strong baselines (1.27 dB PSNR over PGTFormer, 5.6% TLME improvement), and the extension to inpainting and colorization is a nice bonus. The code-aggregation ablation (Top-K vs Average vs theirs) is well designed. Code is open-sourced.\n\nNow the soft spots, in order. Biggest: internal inconsistency. Table 1 lists Ours with TLME=1.091, but Table 3 shows the α=1.0 configuration—same PSNR, AKD, and FVD as Table 1—with TLME=1.107. Meanwhile the \"w/o KL\" row in Table 3 reports TLME=1.091. So as printed, the KL term worsens temporal stability, contradicting the text's claim that ablations validate the Dirichlet prior for temporal coherence. Either Table 1 or Table 3 is wrong, and the takeaway changes depending on which. There's also a smaller arithmetic error in Table 4: the claimed \"5.7% reduction in AKD\" versus Average is actually about 1.6%. These are checkable mistakes.\n\nSecond, the manuscript is incomplete. The degradation pipeline and evaluation metrics are deferred to \"the Appendix,\" but there is no appendix in this arXiv version. The experiments cannot be reproduced from the text. Also, \"differentiably sampling w from a Dirichlet\" is asserted without specifying the reparameterization. The standard location-scale trick doesn't work for Dirichlet; you'd need implicit reparameterization or a score-function estimator, and the gradient estimator determines whether the ELBO is actually being optimized. That is load-bearing for training.\n\nThird, a citation slip: they retrain CodeFormer and cite reference [22], which is a paper on binary code similarity detection, not the CodeFormer face restoration paper (that's [41]). For a method built on top of CodeFormer, this is not a trivial typo.\n\nOne correction to our reader's notes: the KL formula in Eq. (8) is actually fine—ψ(Σ α) is correct for KL(Dir(αhat)||Dir(α))—so the real issue is the table inconsistency, not that formula.\n\nBottom line: the idea is real and the image-quality results are probably in the right ballpark, but the temporal-coherence claim needs rechecking and the missing details must be supplied. Send it to peer review with a request for major revision: provide the appendix, specify the Dirichlet sampling, fix the tables, and correct the citation. Not a desk reject, but not ready as-is.","headline":"Genuinely new idea, inconsistent tables: the paper's temporal-coherence claim is contradicted by its own ablation, so it needs major revision before the numbers can be trusted.","tokens_in":12350,"tokens_out":5888,"would_cite":false,"duration_ms":50484,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A video face restorer trades hard codebook picks for Dirichlet-sampled soft blends to cut flicker.","keywords":["video face restoration","Dirichlet variational inference","codebook learning","temporal coherence","spatio-temporal transformer","blind face restoration","video inpainting","facial colorization"],"falsifier":"Replace the Dirichlet sample in training with the posterior mean (no sampling) and retrain; if PSNR, LPIPS, and TLME stay essentially the same, the reported temporal gains come from the spatio-temporal Transformer or the losses rather than from variational sampling. If they collapse, the unstated differentiable-sampling implementation is the load-bearing detail. A second check: inspect the released code for the sampling operation, since it must be a proper reparameterized Dirichlet draw, not a straight-through or score-function estimator, for the ELBO derivation in the paper to be valid.","tokens_in":11369,"feed_emoji":"🎭","tokens_out":10346,"duration_ms":90095,"temperature":0.7,"pith_summary":"This paper claims that the flicker and temporal instability of codebook-based video face restoration come from hard per-frame discrete codebook lookups, and that replacing those lookups with Dirichlet-distributed soft mixtures of codebook entries removes most of the artifact. The authors build a spatio-temporal Transformer that predicts, for each spatial location and frame, the parameters of a Dirichlet distribution over the codebook; the restored latent code is a convex combination of codebook vectors weighted by a sample from that distribution. Training combines an evidence lower bound (ELBO) over these Dirichlet variables, a Laplacian reconstruction term, and a perceptual (LPIPS) loss. On the public benchmark test set the method reports 29.099 dB PSNR, 0.246 LPIPS, and 1.091 TLME for blind video face restoration, with analogous gains on video face inpainting and colorization. A sympathetic reader would take away that a codebook prior pretrained on still face images can be repurposed for video by making code selection probabilistic rather than discrete.","feed_headline":"Soft codebook blends cut video face flicker at 29.1 dB PSNR","feed_subtitle":"A Dirichlet-soft codebook plus temporal attention restores faces frame by frame with fewer jumps.","key_machinery":"The Dirichlet-constrained codebook is the load-bearing object: it turns the discrete selection of one of N codebook vectors into a distribution over the probability simplex, and each predicted code is a convex combination of codebook entries, so the latent representation lives in the convex hull of the codebook. The spatio-temporal Transformer, alternating spatial and temporal self-attention blocks with sinusoidal positional embeddings, lets the Dirichlet parameters for one frame depend on neighboring frames, which is how smooth transitions between facial features arise. The ELBO objective, with a closed-form Dirichlet KL divergence and a Monte Carlo Laplacian reconstruction term, is the training mechanism that couples the soft blending to reconstruction fidelity.","core_discovery":"The central discovery is that the hard quantization step of a vector-quantized autoencoder can be replaced by a variational Dirichlet relaxation without losing face detail, and that this relaxation is what buys temporal coherence. Concretely, each latent code is written as the weighted average of codebook entries with weights sampled from a Dirichlet distribution whose parameters are predicted by a Transformer that alternates spatial and temporal self-attention over a sliding window of frames. The KL term of the ELBO pushes the predicted Dirichlet parameters toward a chosen prior, so a small prior concentration makes the sampled weights near one-hot while a larger one spreads them out; the reconstruction term, with a Laplacian likelihood and a perceptual loss, keeps the soft codes faithful to face structure. The paper reports that this scheme outperforms per-frame codebook methods and general video restoration baselines across restoration, inpainting, and colorization, and that the temporal metrics improve most.","pith_inferences":["Our inference: the same Dirichlet relaxation should transfer to other video tasks built on discrete codebooks, such as super-resolution, deblurring, or compressed-video enhancement, since nothing in the formulation is face-specific once a suitable codebook prior exists.","Our inference: with the codebook held frozen, the method implicitly separates identity content learned from still images from motion content learned from video; updating the codebook during the video stage is a natural next step and could improve identity retention on very long clips.","Our inference: a direct stress test not reported in the paper is identity-embedding cosine similarity across restored frames; the reported landmark and attribute metrics suggest it would improve, but that measurement would settle whether temporal coherence is actually identity preservation rather than just low pixel jitter."],"forward_implications":["Any vector-quantized autoencoder pretrained on high-quality still images can be adapted to video by swapping hard quantization for Dirichlet-soft blending, with no need to retrain the codebook from scratch.","One spatio-temporal framework covers restoration, inpainting, and colorization, and the reported drop in the temporal flicker metric shows that frame-to-frame jumps are reduced.","The sliding-window, center-frame inference scheme keeps memory bounded for arbitrary-length videos, since only a short window needs to be processed at a time.","Because every restored latent code carries an explicit probability distribution over codebook entries, the model yields a per-region uncertainty signal that could flag unreliable reconstructions."],"supporting_citations":[{"why":"Supplies the codebook-pretraining paradigm that the paper extends from still images to video.","marker":"[8]"},{"why":"Defines the codebook-lookup restoration baseline whose pretrained weights initialize the model and which serves as a comparison.","marker":"[41]"},{"why":"Provides the video face dataset, degradation pipeline, and test split used for training and evaluation.","marker":"[32]"},{"why":"Is the parsing-guided temporal-coherent transformer baseline that the paper compares against.","marker":"[33]"},{"why":"Is a generative facial-prior restoration baseline used for comparison on the restoration task.","marker":"[30]"},{"why":"Is a GAN-based restoration baseline used for comparison on restoration, inpainting, and colorization.","marker":"[35]"},{"why":"Is a video face super-resolution baseline included in the main comparison table.","marker":"[9]"},{"why":"Supplies the perceptual metric used both as a training loss and as an evaluation measure.","marker":"[39]"}],"fun_headline_variants":["Dirichlet soft codebook cuts video face flicker","Temporal face restoration via Dirichlet codebook","Soft codebook keeps video faces stable","Dirichlet relaxation stabilizes face videos","Variational codebook for flicker-free faces"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The training procedure assumes that samples from the Dirichlet posterior can be drawn 'differentiably' in Eq. (10), but the paper never states the reparameterization; without a correct gradient estimator for Dirichlet sampling, the ELBO gradients and therefore the reported gains are not actually guaranteed by the argument given.","fun_headline_variants_meta":{"raw":{"variants":["Dirichlet soft codebook cuts video face flicker","Temporal face restoration via Dirichlet codebook","Soft codebook keeps video faces stable","Dirichlet relaxation stabilizes face videos","Variational codebook for flicker-free faces"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00024,"raw_usage":{"total_tokens":1509,"prompt_tokens":928,"completion_tokens":581,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":544,"completion_tokens_details":{"reasoning_tokens":513}},"tokens_in":544,"tokens_out":581,"duration_ms":5205,"temperature":1.0,"reasoning_tokens":513,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T20:04:14.478208+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Replace the Dirichlet sample in training with the posterior mean (no sampling) and retrain; if PSNR, LPIPS, and TLME stay essentially the same, the reported temporal gains come from the spatio-temporal Transformer or the losses rather than from variational sampling. If they collapse, the unstated differentiable-sampling implementation is the load-bearing detail. A second check: inspect the released code for the sampling operation, since it must be a proper reparameterized Dirichlet draw, not a straight-through or score-function estimator, for the ELBO derivation in the paper to be valid.","supporting_citations":[{"cited_title":"Taming transformers for high-resolution image synthesis","cited_arxiv_id":null,"evidence_quote":"Supplies the codebook-pretraining paradigm that the paper extends from still images to video."},{"cited_title":"Towards robust blind face restora- tion with codebook lookup transformer","cited_arxiv_id":null,"evidence_quote":"Defines the codebook-lookup restoration baseline whose pretrained weights initialize the model and which serves as a comparison."},{"cited_title":"Vfhq: A high-quality dataset and bench- mark for video face super-resolution","cited_arxiv_id":null,"evidence_quote":"Provides the video face dataset, degradation pipeline, and test split used for training and evaluation."},{"cited_title":"Beyond alignment: blind video face restoration via parsing- guided temporal-coherent transformer","cited_arxiv_id":null,"evidence_quote":"Is the parsing-guided temporal-coherent transformer baseline that the paper compares against."},{"cited_title":"Gan prior embedded network for blind face restoration in the wild","cited_arxiv_id":null,"evidence_quote":"Is a GAN-based restoration baseline used for comparison on restoration, inpainting, and colorization."},{"cited_title":"Kalman-inspired feature propagation for video face super- resolution","cited_arxiv_id":null,"evidence_quote":"Is a video face super-resolution baseline included in the main comparison table."},{"cited_title":"The unreasonable effectiveness of deep features as a perceptual metric","cited_arxiv_id":null,"evidence_quote":"Supplies the perceptual metric used both as a training loss and as an evaluation measure."}],"review_version":1}