{"id":"2b7abfc4-2183-4b3f-8ea6-e99116089a2f","arxiv_id":"2601.19506","paper_version":4,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"high","formal_verification":"none","parameter_count":6,"one_line_summary":"Pref-Restore combines AR semantic tokens, a diffusion generator, and DiffusionNFT-style RL to make blind face restoration more consistent, but its deterministic-identity claim is weakened by self-referential rewards and identity loss in the quality variant.","lead":"This paper builds a face-restoration system that combines an auto-regressive language model, a diffusion generator, and reinforcement learning, claiming more identity-consistent and less variable outputs. Why read: if it holds up, it would make restored faces more trustworthy for identity-sensitive uses and shows a way to inject preference feedback into diffusion restoration.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Deterministic claim is self-referential: variance is measured on the same reward scores optimized in Eq. 15, and Stage-2 Quality degrades identity (ArcFace Deg 54.06→73.99), so 'lower uncertainty' may be reward-mode collapse, not faithful restoration.","rationale":"The paper's central claim is deterministic, identity-consistent blind face restoration. The strongest evidence for determinism is Sec. V-H's variance reduction in HPSv2/PickScore/ClipScore/Sum Score. This evidence is structurally circular: those scores are the components of R_pref in Eq. (15), so a policy that maximizes expected reward will typically also reduce reward variance over samples, regardless of whether the posterior over the correct identity is sharpened. A model that collapses to a single over-smoothed or beautified output would show the same '50% reduction.' The paper's own Table II shows the Stage-2 Quality variant—the one receiving the RL fine-tuning—suffers a major identity-fidelity regression (ArcFace Deg 54.06→73.99, LMD 5.13→8.36), while the Fidelity variant with the best identity metrics is only Stage 1 and was not the subject of the deterministic analysis in Fig. 8. Appendix I(b) further concedes that the current reward is primarily aesthetic and that identity-preserving rewards are future work. No code or checkpoints are released to verify the stability protocol or the independence of the seeds. The Stage-1.2 ablation (Table IX) does support texture alignment as a legitimate source of identity gains, so the architecture has a plausible mechanism; the unsupported part is the additional claim that preference RL makes restoration deterministic without sacrificing identity. The proposed concrete test—measuring identity-metric variance and mean across seeds, plus using an independent judge—would directly settle whether the concern lands.","tokens_in":31831,"tokens_out":4054,"duration_ms":50140,"concrete_test":"Re-run the Sec. V-H N=16 stability protocol on the same test images, but compute per-image standard deviation and mean of identity-sensitive metrics (ArcFace embedding cosine/angle, LMD) for Base, Pref-Restore Fidelity, and Pref-Restore Quality, both with and without captions. If the Quality variant's identity-metric std is not significantly lower than Base, or its mean identity is significantly worse, the deterministic claim must be narrowed to 'reduced variance in the aesthetic reward used for training.' Additionally, score all variants with an independent held-out preference judge not used in R_pref (e.g., human pairwise ratings or a different IQA model) to rule out reward overfitting as the source of the variance drop.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central assertion 'deterministic, identity-consistent restoration' rests on Sec. V-H, where the median std of the Sum Score drops ~50% across N=16 repeated restorations. But that Sum Score is HPSv2 + PickScore + ClipScore—exactly the reward R_pref optimized in Eq. (15). Reducing variance in the training objective is evidence of convergence to a reward-maximizing mode, not of reduced posterior entropy over the true identity; it could equally arise from mode collapse or from consistently producing aesthetically pleasing but identity-wrong faces. The paper's own data contradict the identity part: Table II shows the Stage-2 Quality variant—the one after preference-aware fine-tuning—increases ArcFace Deg from 54.06 to 73.99 and LMD from 5.13 to 8.36, i.e., it is substantially less identity-faithful than the Stage-1 Fidelity variant. The Fidelity variant's identity gains are attributable to Stage 1.2 texture alignment (Table IX), not to the preference RL that is claimed to prune stochasticity. Appendix I(b) reinforces the gap: the authors state they 'primarily utilized an aesthetic reward model' and list identity-preserving rewards as future work. Thus the evidence supports 'preference-aligned aesthetic stabilization,' not 'deterministic blind face restoration with identity-sensitive fidelity.'","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Pref-Restore, a hierarchical blind face restoration framework combining an auto-regressive semantic integrator (image + text to discrete tokens) with a diffusion-based generator, trained in two stages: Stage 1 performs semantic-to-diffusion and texture-to-diffusion alignment, and Stage 2 fine-tunes the diffusion velocity field via DiffusionNFT-style on-policy RL with a composite reward (HPSv2 + Clip Score + Pick Score). The authors report SOTA results on synthetic and real-world benchmarks and, as the central claim, assert that Stage 2 reduces restoration stochasticity by about 50% (Sec. V-H, Fig. 8) and yields deterministic, identity-consistent restoration. Two variants are offered: Pref-Restore Fidelity (after Stage 1) and Pref-Restore Quality (after Stage 2).","tokens_in":32234,"tokens_out":4263,"duration_ms":54577,"significance":"If the deterministic-restoration claim were fully supported, the paper would make a valuable contribution: using AR-derived semantic anchors and preference-based RL to stabilize diffusion restoration is a plausible route toward mitigating hallucination and identity drift in blind face restoration. The paper deserves credit for its extensive ablations (Table IX, Fig. 3, Table VII), the staged training design, and the decision to provide two clearly characterized variants. However, the central claim is not currently supported: the variance reduction is measured on the very reward function being optimized, and the authors' own Table II shows the Stage 2 'Quality' variant degrading identity metrics substantially. As written, the evidence supports 'preference-aligned aesthetic stabilization' rather than 'deterministic, identity-consistent restoration.' A major revision is needed to reframe the claim or add evidence from identity-based uncertainty metrics.","major_comments":[{"comment":"The '50% reduction in restoration stochasticity' is measured as the standard deviation of the Sum Score (HPSv2 + PickScore + ClipScore) over 16 runs. This is exactly the reward R_pref optimized in Eq. (15) and tracked in Fig. 6. Reduced variance of the training objective across seeds is expected when the model converges to a reward-maximizing mode, and it can also result from mode collapse. It does not measure uncertainty over the restored identity or output distribution. The paper needs independent uncertainty metrics (e.g., std of ArcFace embeddings, landmark positions, or pixel-space distance between restorations) before the determinism claim can be accepted.","section":"Sec. V-H and Eq. (15)"},{"comment":"The abstract claims 'stronger identity-sensitive fidelity' and the Method section claims Stage 2 'prunes trajectories that deviate from human aesthetic and fidelity standards.' Yet Table II shows Pref-Restore Quality (the Stage 2 variant) has LMD 8.3588 and ArcFace Deg 73.9882, much worse than Pref-Restore Fidelity (5.1337 / 54.0623), and worse than several baselines. Table IX attributes the identity gains to Stage 1.2, not to preference RL. Thus the stage that produces the claimed stochasticity reduction is also the stage that degrades identity. The deterministic-identity reading is contradicted by the authors' own data; the claim should be restricted to aesthetic stabilization, or Stage 2 must be shown to preserve identity via an identity-aware reward.","section":"Table II and Sec. V-B/G"},{"comment":"On the four real-world datasets, the reported metrics are almost entirely no-reference aesthetic/quality metrics (MUSIQ, CLIPIQA+, MANIQA, topiq, DSL-FIQA) plus FID. Identity-sensitive metrics (ArcFace Deg, LMD) that appear in the synthetic CelebA-Test table are omitted. Since the paper's headline contribution is 'identity-sensitive fidelity' and 'lower restoration uncertainty,' real-world identity metrics (or a clear explanation for their omission) are necessary to support the generalization claim.","section":"Tables III-VI"},{"comment":"The text reporting reward gains is internally inconsistent: 'the Pick Score rises from 76.98 to 86.66 (+12.6%)' while Table VIII lists Pick Score 20.01 to 22.53, and HPSv2 is listed as 22.28 to 30.22 in the table but appears in a different scale in the text. If a rescaling or normalization is used, it must be stated explicitly; otherwise the main quantitative evidence for Stage 2's benefit is not verifiable.","section":"Sec. V-E and Table VIII"}],"minor_comments":[{"comment":"The formal derivation in Appendix A is a standard MAP-plus-energy-based-prior rewrite; Eq. (7) is the elementary 'conditioning reduces entropy' inequality. The paper should not present this as a new theoretical proof. The actual contribution is the empirical architecture, which is fine, but the framing should be modest.","section":"Eq. (6) and Appendix A"},{"comment":"The values of gamma, lambda, Z, and the reward composition weights are not reported. Table I lists learning rates and batch size but not these Stage-2-specific hyperparameters. Please provide them for reproducibility.","section":"Table I / Eqs. (12)-(14)"},{"comment":"The figure and text do not specify which model is the 'Base Model' (presumably Stage 1?) and how the Sum Score is computed/scaled. The boxplot axes are unlabeled beyond 'Std of Scores'; clarify the normalization of HPSv2/PickScore/ClipScore in this analysis.","section":"Fig. 8"},{"comment":"The caption reads 'Wilder-Test' but the dataset is WIDER-Test. Fix the typo.","section":"Table V caption"},{"comment":"Given DiffusionReward [46], IRPO [49], RealSR-R1 [50], and TTPO [51] already apply RL/preference methods to restoration, the claim to 'pioneer' on-policy RL for BFR should be softened or carefully differentiated from these prior works.","section":"Related Work / Contributions"}],"recommendation":"major_revision","confidential_remarks":"The paper's engineering is solid and the Stage 1 results (Fidelity variant) are competitive, but the central determinism/identity claim is not supported by the current evidence. The most efficient fix is to reframe the paper around 'preference-aligned aesthetic stabilization' and report uncertainty in identity-sensitive metrics; alternatively, add an identity-preserving reward and rerun the stochasticity analysis. I recommend major revision rather than rejection because the underlying framework is promising and the load-bearing issue is fixable within the manuscript's scope."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Candid take: this is a competent systems paper with an unsupported headline. The genuinely new part is applying DiffusionNFT-style forward-flow RL to blind face restoration, on top of an AR semantic tokenizer feeding a conditional flow-matching generator. That specific combination I have not seen before, and the two-variant design (Fidelity vs. Quality) is a sensible way to expose the trade-off. Credit where due: the ablation in Table IX shows Stage 1.2 texture alignment gives a large real gain (Deg 74.5→54.1), and the authors are honest in Sec. V-G and Appendix I(b) that their RL stage mostly optimizes aesthetics and that identity-preserving rewards are future work.\n\nThe soft spots are load-bearing. The 'deterministic restoration' claim in Sec. V-H is measured as lower variance of HPSv2/PickScore/ClipScore across 16 runs. Those are exactly the components of R_pref in Eq. (15). Lower variance in the training objective is evidence of convergence to a reward-maximizing mode, not lower posterior entropy over the true identity. Their own Table II makes the identity problem explicit: the Quality variant (after RL) goes from Deg 54.06 to 73.99 and LMD 5.13 to 8.36, so it is substantially less faithful than the Stage 1 model. The abstract's 'identity-consistent deterministic' phrasing is contradicted by their Table II. The Fidelity variant's identity gains come from Stage 1.2, not from the preference RL.\n\nTwo more issues. The SOTA comparison is confounded: their model sees Qwen-generated captions, and none of the baselines get text. They report a caption dropout in Appendix F, but no no-caption evaluation or text-guided baseline comparison. Also Table VIII has a concrete inconsistency: the text says Pick Score rises from 76.98 to 86.66, but the table shows 20.01 to 22.53. That must be fixed. No code or checkpoints is a real problem for a paper whose central claim is reduced stochasticity.\n\nThe formal derivation in App. A is a standard MAP-plus-energy-prior rewrite, not a new proof; fine as framing, not a contribution. The framework itself is plausible, the ablations are informative, and the authors do acknowledge the fidelity/quality trade-off. But the headline claims need to be reframed: either 'lower variance in preference scores' or, better, independent held-out judges, identity metrics on real benchmarks, no-caption comparisons, and a corrected Table VIII.\n\nRecommendation: send to peer review — this deserves referee time, not a desk reject — but expect major revision. The reviewer should push on the deterministic claim and the caption confound. A serious referee could help turn this into a solid paper with a defensible conclusion.","headline":"The RL-plus-AR-token system is a plausible engineering combination, but the headline 'deterministic, identity-consistent' claim rests on variance measured on the very rewards they optimized, and their own Table II shows the RL-tuned variant loses identity.","tokens_in":32720,"tokens_out":1958,"would_cite":false,"duration_ms":25675,"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":"Blind face restoration can be made deterministic by anchoring a diffusion generator with semantic tokens and pruning its output space with preference-aware reinforcement learning.","keywords":["Blind face restoration","Information asymmetry","Autoregressive models","Diffusion models","On-policy reinforcement learning","Preference alignment","Deterministic restoration","Identity preservation"],"falsifier":"Run N=16 restorations of the same degraded face under different random seeds and compute the standard deviation of identity metrics (face-embedding angle and landmark distance) rather than preference scores. If those spreads do not shrink alongside the preference-score spread—or if the mean identity error grows, as the paper's own Table II already shows for stage 2—then the claimed determinism is a property of the reward metric, not of the restored identity.","tokens_in":31699,"feed_emoji":"👤","tokens_out":7245,"duration_ms":79735,"temperature":0.7,"pith_summary":"The paper argues that blind face restoration fails not because generators are weak but because the task is information-asymmetric: a severely degraded face can map to many plausible but identity-wrong faces. It proposes Pref-Restore, which attacks both ends of the pipeline—an autoregressive semantic branch turns image and text cues into discrete tokens that anchor the high-level structure, while a preference-aware reinforcement-learning stage prunes the diffusion output distribution toward human-aligned, identity-consistent results. The central claim is that this two-ended design makes restoration effectively deterministic, cutting the median standard deviation of preference scores across repeated sampling by about 50 percent while matching or beating prior methods on perceptual and face-specific metrics. A sympathetic reader would care because deterministic, identity-faithful restoration is what forensic, medical, and archival applications require.","feed_headline":"Pref-Restore cuts face restoration randomness in half","feed_subtitle":"Autoregressive semantic tokens anchor the diffusion output; preference-aware RL prunes the solution space.","key_machinery":"Three coordinated components carry the argument. (1) An AR-based semantic integrator: a quantized visual-language model trained by next-token prediction turns degraded pixels plus a text caption into discrete semantic tokens that serve as a stable high-level anchor. (2) A continuous diffusion generator built on conditional flow matching, trained in two sub-stages (semantic-to-diffusion alignment, then texture-to-diffusion alignment) so the anchor and low-level VAE features jointly condition the velocity field. (3) Preference-aware fine-tuning via an on-policy RL scheme applied to the forward diffusion flow (the paper calls this DiffusionNFT): group rollouts produce K candidates scored by a f","core_discovery":"On its own terms, the paper's discovery is that the ill-posedness of blind face restoration can be re-balanced rather than merely regularized. Pref-Restore augments the degraded input with dense semantic tokens generated by an autoregressive visual-language model, then uses on-policy reinforcement learning on the forward diffusion flow to make the generator prefer high-fidelity trajectories and avoid hallucinated ones. The paper reports that the resulting model matches or exceeds prior methods on synthetic and real benchmarks, with stronger identity-sensitive fidelity after stage one and higher perceptual quality after stage two, and that across 16 repeated restorations the spread of prefere","pith_inferences":["A stronger test of 'deterministic restoration' than the paper reports would measure the spread of identity metrics (face-embedding angle and landmark distance) across the same 16 runs; if those spreads do not tighten alongside the preference-score spread, the 50% reduction may describe reward collapse rather than faithful reconstruction.","The paper's own Table II shows stage 2 degrades identity metrics (embedding angle from 54.06 to 73.99), so 'preference-aligned' currently means aesthetically preferred, not identity-faithful; adding an explicit identity-reward term would be the natural next step.","Because captions are generated by a separate visual-language model in deployment, caption errors become a new failure mode: a wrong attribute in the text could anchor the restoration to the wrong identity. A natural stress test is corrupting or dropping captions and measuring identity fidelity and determinism.","The recipe—autoregressive semantic anchoring plus preference-aware forward-flow RL—generalizes beyond faces to other ill-posed inverse problems such as blind super-resolution, deblurring, or medical image restoration where structure matters more than aesthetic plausibility."],"forward_implications":["If the central claim is right, a single degraded face will restore to the same identity-consistent output across repeated sampling runs, which is what forensic, archival, and medical uses demand.","Text captions act as a steerable semantic channel: richer captions improve semantic fidelity (the paper reports double-digit gains in CLIP-image similarity and DreamSim), so the same restoration engine can be directed by language at inference time.","The two variants make the fidelity–quality trade-off explicit and controllable: stopping after stage 1 preserves identity best, while stage 2 maximizes perceptual scores but measurably worsens identity metrics—so users can choose where to sit on the curve.","Because the RL stage needs only about 30 fine-tuning steps and operates on the forward velocity field rather than the reverse sampling policy, the same distribution-pruning recipe can be attached to other diffusion-based restoration models without re-deriving policy gradients."],"fun_headline_variants":["Pref-Restore anchors face restoration with semantic tokens","RL refines face restoration to cut identity errors","Deterministic face restoration via semantic anchoring","Pref-Restore reduces hallucination in face restoration","Semantic anchors boost face fidelity in blind restoration"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The load-bearing premise is that the drop in variance of preference scores (human-preference, CLIP-alignment, and aesthetic-quality rewards) across repeated sampling measures a genuine reduction in restoration uncertainty; if the reward model simply makes the model consistently produce a pleasing but wrong face, the deterministic claim collapses.","fun_headline_variants_meta":{"raw":{"variants":["Pref-Restore anchors face restoration with semantic tokens","RL refines face restoration to cut identity errors","Deterministic face restoration via semantic anchoring","Pref-Restore reduces hallucination in face restoration","Semantic anchors boost face fidelity in blind restoration"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000477,"raw_usage":{"total_tokens":2201,"prompt_tokens":746,"completion_tokens":1455,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":490,"completion_tokens_details":{"reasoning_tokens":1383}},"tokens_in":490,"tokens_out":1455,"duration_ms":11403,"temperature":1.0,"reasoning_tokens":1383,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-03T07:33:41.794427+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run N=16 restorations of the same degraded face under different random seeds and compute the standard deviation of identity metrics (face-embedding angle and landmark distance) rather than preference scores. If those spreads do not shrink alongside the preference-score spread—or if the mean identity error grows, as the paper's own Table II already shows for stage 2—then the claimed determinism is a property of the reward metric, not of the restored identity.","supporting_citations":[],"review_version":1}