{"id":"cd0d51fc-45d7-4126-9784-d7a33b6fb2ac","arxiv_id":"1908.08239","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"A progressive face super-resolution network with a landmark-heatmap-weighted attention loss produces 8x upscaled faces that human raters judged more realistic than prior methods.","lead":"The authors train a face super-resolution network that enlarges 16x16 face crops to 128x128 in three progressive steps, using landmark heatmaps to focus the loss on eyes, nose, and mouth. The method reports better perceptual quality than prior face SR systems, while a distilled landmark detector cuts training time from about 3 days to 1 day.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Abstract claims superiority in 'both qualitative and quantitative measurements,' but Table 3 shows VDSR wins PSNR on both datasets; the quantitative claim needs restriction.","rationale":"The reader's verdict is CONDITIONAL, and the reader's rationale already flags the VDSR PSNR issue and the absence of confidence intervals, so our concern reinforces the existing verdict rather than moving it. We chose the quantitative-claim contradiction as the most load-bearing issue because it directly bears on the paper's strongest stated claim. The reader's weakest_assumption focused on the distilled FAN heatmaps; that issue is at least partially addressed by the paper's own comparison in Table 1 between the original FAN and the distilled FAN, although without error bars. The missing loss weights in Eq. 6 and the apparent miscitation of WGAN-GP are additional reproducibility issues, but they are secondary to the claim-evaluation mismatch. Overall, the paper contains a plausible method and an internal ablation, but the central claim needs to be narrowed to the metrics where the method actually wins.","tokens_in":8631,"tokens_out":6806,"duration_ms":73605,"concrete_test":"Run the released checkpoints on the same test splits and compute per-image PSNR, SSIM, and MS-SSIM with 95% confidence intervals and paired significance tests against VDSR and FSRGAN. Separately obtain the MOS raw scores and report the number of rated scenes per dataset, per-rater and per-scene variance, and 95% confidence intervals for Ours versus each baseline. If VDSR remains significantly higher in PSNR on either dataset, the abstract must be revised to claim superiority in structural and perceptual quality, not in 'both qualitative and quantitative measurements.'","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's strongest claim is that the method 'outperforms state-of-the-art methods in both qualitative and quantitative measurements, especially in perceptual quality.' The paper's own quantitative comparison in Table 3 contradicts this as stated: VDSR achieves higher PSNR than the proposed method on both aligned (22.94 vs. 22.66) and unaligned (23.70 vs. 22.96) test sets. PSNR is a quantitative measurement, so the unqualified claim of superiority in quantitative measurements is not supported. The text acknowledges that VDSR wins PSNR but dismisses the result because VDSR outputs are blurred; this is reasonable for a perceptual-quality argument, but it means the abstract should say the method wins on structural and perceptual metrics (SSIM, MS-SSIM, MOS) while conceding PSNR. Additionally, no confidence intervals or significance tests are reported for any metric, and the MOS section omits the number of rated test scenes, per-rater variance, and confidence intervals, so the magnitude of the claimed perceptual advantage cannot be independently assessed. The central method may still be promising, but the central claim as worded is overbroad and should be narrowed.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a progressive face super-resolution method that upscales faces 8x using a generator trained in three steps, a facial attention loss that weights pixel differences by landmark heatmaps, and a distilled (compressed) Face Alignment Network used to produce those heatmaps. The method is evaluated on aligned and unaligned CelebA and AFLW test sets with PSNR, SSIM, MS-SSIM, and a MOS perceptual test, and the authors report state-of-the-art results especially in perceptual quality, with a compact landmark network and reduced training time.","tokens_in":8851,"tokens_out":3361,"duration_ms":30907,"significance":"If the reported results hold, the paper offers a useful combination of progressive training, landmark-guided attention, and model distillation for face super-resolution. The ablation study in Table 2 supports the contribution of the facial attention loss and progressive training, and the GitHub release makes the approach reproducible. The main value is the perceptual-quality improvement shown in the MOS test, though that test currently lacks the statistical detail needed to verify the strength of the claim.","major_comments":[{"comment":"The abstract's claim that the method 'outperforms state-of-the-art methods in both qualitative and quantitative measurements' is not supported by the paper's own Table 3: VDSR reports higher PSNR on both the aligned (22.94 vs. 22.66) and unaligned (23.70 vs. 22.96) test sets, and PSNR is a quantitative measurement. The text's explanation that VDSR outputs are blurred is reasonable, but the quantitative claim must be narrowed to structural and perceptual metrics (SSIM, MS-SSIM, MOS) with an explicit acknowledgment that the method does not lead on PSNR.","section":"Abstract; Table 3"},{"comment":"The MOS test is the key evidence for the perceptual-quality claim, but no confidence intervals, per-rater variance, or significance tests are reported, and the paper refers to 'Supplementary Materials' for details that are not included in the submission. Without these, the statement that the method shows 'overwhelming performance' in Figure 7 cannot be assessed; at minimum the authors should report the number of scenes rated, inter-rater agreement, and a paired significance test against the closest competitor.","section":"Section 4.4 (MOS)"},{"comment":"The claim that the distilled FAN produces heatmaps 'suitable for facial attention weights' is validated only indirectly: Table 1 reports NME computed against the original FAN's predictions, which measures agreement with the teacher rather than landmark accuracy, and the PSNR/SSIM improvements attributed to the distilled FAN are small and given without variance estimates. The authors should provide an independent landmark-localization evaluation (e.g., against manual annotations) and error bars, or explicitly restrict the claim to 'agrees with the original FAN while improving SR metrics in our pipeline.'","section":"Section 4.2; Table 1"}],"minor_comments":[{"comment":"Table 1 contains malformed entries (e.g., '(64x64)- (128x128)-', '0.830% 2.2643%22.96') and the column labels are garbled; please reformat the table so each metric is a separate numeric column.","section":"Table 1"},{"comment":"Equation (1) is defined with G(ILR), but the loss is applied at Steps 2 and 3 with different upscale factors; please clarify which step's generator output is used in each application.","section":"Section 3.2"},{"comment":"The normalization of M is unclear: the text says the heatmap is 'min-max normalized into [0,1]' and that M* is the 'channel-wise max values'; please specify whether normalization is applied per landmark channel or globally across all channels.","section":"Section 3.2"},{"comment":"The notation N∈{2,3} is used for the step number, but N is also used for the number of landmarks in Eq. (5); please use distinct symbols for these two quantities.","section":"Section 3.1"},{"comment":"The overall loss in Eq. (6) depends on α, β, γ, λ, and η, but their values are never reported; please include them in the implementation details.","section":"Section 4.1; Eq. (6)"},{"comment":"The gradient penalty is attributed to reference [23], but WGAN-GP is normally credited to Gulrajani et al. (NIPS 2017); please verify the citation.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The paper is clearly written and the code release is a plus, but the central claim of quantitative superiority is contradicted by the paper's own PSNR numbers, and the MOS evidence is reported without statistical support. I recommend major revision with a focus on recalibrating the claims and adding the missing statistical details; the core method appears defensible."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper is a competent, incremental contribution to face super-resolution. The new pieces are a heatmap-weighted L1 attention loss, progressive training applied to face SR, and a distilled FAN that cuts the landmark network to about 1% of the original parameters. The combination is new, the writing is clear, and the ablation study shows each component earns its place. The MOS result is striking: 3.73 versus the next best 2.55, which is a large perceptual gap. Code is available, which helps reproducibility. Credit where it is due: the authors did the experiments that back the core claim, and the distilled FAN idea is genuinely useful for anyone who wants cheap landmark heatmaps in a GAN pipeline. The soft spots are real but addressable. The abstract says the method outperforms state-of-the-art in both qualitative and quantitative measurements, yet Table 3 shows VDSR wins PSNR on both datasets. The authors even acknowledge this, which is fine, but the abstract should be narrowed to structural and perceptual metrics. The stress-test note is correct here. More broadly, no metric has confidence intervals or significance tests, and the MOS section omits per-rater variance and the number of rated scenes, so we cannot judge whether the perceptual gap is robust. The loss weights in Eq. 6 are unspecified, which is annoying for reproducibility. The claim that the distilled FAN gives suitable heatmaps is mostly validated inside the authors' own pipeline, though the NME comparison against the original FAN does provide an independent anchor; it is not a fabricated metric. These are minor-to-moderate flaws, not load-bearing ones. The reader's conditional verdict matches mine. This paper is not a revolution, but it is a plausible and useful step for a narrow subfield. I would send it to peer review; with a revised abstract and error bars, it could be a solid conference paper. For a reading group, it is worth a look if you work on face SR or GAN-based perceptual losses.","headline":"A solid incremental face SR paper whose perceptual gains look real, but the abstract overclaims quantitative superiority and the missing error bars need fixing before publication.","tokens_in":603,"tokens_out":901,"would_cite":true,"duration_ms":23379,"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":"Landmark heatmaps, used as attention weights, let a progressive network super-resolve faces 8x with photo-realistic detail.","keywords":["face super-resolution","progressive training","facial landmark heatmap","attention loss","network distillation","generative adversarial network","perceptual quality","8x upscaling"],"falsifier":"Retrain the same progressive network with the attention heatmaps deliberately corrupted—shifted by several pixels, replaced by random maps, or swapped between paired images—and compare SSIM and MOS on the test set; if the reported gains do not clearly disappear, the facial attention loss is not responsible for the claimed detail restoration.","tokens_in":8433,"feed_emoji":"🎯","tokens_out":7209,"duration_ms":62777,"temperature":0.7,"pith_summary":"The paper sets out to show that face super-resolution in the extreme 8x regime can be made stable and perceptually realistic by training in progressive 2x steps and by steering the reconstruction loss with facial landmark heatmaps. The network is claimed to restore fine details around eyes, nose, and mouth without distortion, and to beat prior face SR methods on structural similarity and human perceptual ratings, even though its pixel PSNR is slightly below a very deep non-face baseline. The authors argue this works because the heatmaps come from a lightweight distilled face alignment network whose responses concentrate on non-occluded landmark areas.","feed_headline":"Landmark heatmaps steer 8x face super-resolution to sharper details","feed_subtitle":"The network upscales 16-pixel faces to 128 pixels while landmark heatmaps keep eyes, nose, and mouth from blurring.","key_machinery":"The load-bearing mechanism is the facial attention loss, defined as $L_{\\text{attention}} = \\frac{1}{r^2 WH}\\sum_{x,y} M^*_{x,y}\\, |I^{HR}_{x,y} - G(I^{LR})_{x,y}|$, where $M^*$ is the channel-wise maximum of the target image's landmark heatmap after min-max normalization to [0,1]. This loss is applied at the 4x and 8x steps of a progressively grown generator–discriminator pair that uses residual blocks, transpose convolutions, and a Wasserstein adversarial loss with gradient penalty. The heatmaps come from a distilled face alignment network, trained by hint-based regression to match the original FAN's heatmaps; because it lacks the encoder–decoder and skip-layer structure, its responses are broad over visible landmark regions and low over occluded ones, which the paper argues makes them suitable attention weights. The machinery also includes a heatmap loss that penalizes disagreement between heatmaps of the generated and target images to preserve structural consistency.","core_discovery":"On the paper's own terms, the central discovery is that an 8x face SR generator can be trained progressively—growing from 2x to 4x to 8x—and that at each later stage a facial attention loss should multiply the L1 pixel difference by a min-max-normalized landmark heatmap. The heatmap is the channel-wise max of the heatmaps produced by a distilled face alignment network, so the loss focuses the generator on regions around facial landmarks while suppressing occluded regions. A corresponding heatmap loss keeps the structure of the generated face aligned with the target. The result is a network that produces 128x128 images from 16x16 inputs with preserved facial attributes, and the authors report that it outperforms prior methods on SSIM, MS-SSIM, and a mean-opinion-score perceptual test.","pith_inferences":["Editorial inference: the same heatmap-weighted loss could be applied to other landmark-structured objects, such as hands, bodies, or medical structures, whenever a cheap landmark estimator is available, making the attention mechanism task-agnostic.","Editorial inference: the distilled FAN's low response on occluded landmarks suggests the attention loss doubles as an occlusion-aware weighting; a direct test would compare aligned versus misaligned heatmap weights on occluded faces.","Editorial inference: because PSNR is slightly lower than VDSR while MOS is much higher, the metric story implies pixel-level fidelity and human-perceived quality diverge more sharply for faces than for general images; this could be tested by ablating the adversarial loss while keeping the attention loss."],"forward_implications":["8x face super-resolution can be trained stably without attribute vectors or other priors that are hard to obtain in the wild.","Applying the landmark attention loss at intermediate progressive steps yields measurable gains in SSIM, MS-SSIM, and human MOS over applying the same losses without progressive training.","A distilled landmark network with roughly 1.35% of the original FAN's parameters produces heatmaps that work better for attention and cuts training time from about 3 days to about 1 day.","The method generalizes to both aligned and unaligned face crops, suggesting it does not depend on tight face pre-registration.","The reported MOS gap over prior face SR methods implies the main practical advantage is perceptual realism rather than raw pixel accuracy."],"supporting_citations":[{"why":"Supplies the original FAN whose heatmaps the distilled FAN imitates and compares against.","marker":"[4]"},{"why":"Provides the hint-based distillation objective used to train the compressed FAN.","marker":"[17]"},{"why":"Defines FSRNet and FSRGAN, face SR baselines that use facial priors and are compared in the experiments.","marker":"[5]"},{"why":"Motivates progressive growing of generator and discriminator, the training schedule adopted by the paper.","marker":"[19]"},{"why":"Delivers the WGAN gradient-penalty term that stabilizes the adversarial training.","marker":"[23]"},{"why":"Introduces the heatmap loss for aligning landmarks in super-resolved faces, which the paper adds to its objective.","marker":"[1]"},{"why":"Defines the perceptual loss on deep features used to avoid blurry outputs.","marker":"[7]"},{"why":"Supplies VDSR, the very deep super-resolution baseline that achieves the highest PSNR but blurry results in the comparison.","marker":"[8]"},{"why":"Supplies URDGN, a GAN-based face super-resolution baseline used for quantitative and MOS comparison.","marker":"[24]"}],"fun_headline_variants":["Landmark heatmaps guide progressive 8x face upscaling","Sharp 8x faces via progressive training and landmark attention","Progressive face SR that focuses on facial landmarks for 8x detail","8x super-resolution with progressive steps and heatmap focus","Attention to landmarks yields sharper progressive 8x face SR"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the distilled landmark heatmaps reliably mark which pixels carry facial detail worth preserving, so that weighting the reconstruction loss by those maps improves actual facial fidelity rather than merely sharpening landmark regions.","fun_headline_variants_meta":{"raw":{"variants":["Landmark heatmaps guide progressive 8x face upscaling","Sharp 8x faces via progressive training and landmark attention","Progressive face SR that focuses on facial landmarks for 8x detail","8x super-resolution with progressive steps and heatmap focus","Attention to landmarks yields sharper progressive 8x face SR"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000217,"raw_usage":{"total_tokens":1412,"prompt_tokens":895,"completion_tokens":517,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":511,"completion_tokens_details":{"reasoning_tokens":431}},"tokens_in":511,"tokens_out":517,"duration_ms":5401,"temperature":1.0,"reasoning_tokens":431,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T11:44:48.608624+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Retrain the same progressive network with the attention heatmaps deliberately corrupted—shifted by several pixels, replaced by random maps, or swapped between paired images—and compare SSIM and MOS on the test set; if the reported gains do not clearly disappear, the facial attention loss is not responsible for the claimed detail restoration.","supporting_citations":[{"cited_title":"How far are we from solving the 2d & 3d face align- ment problem? (and a dataset of 230,000 3d facial landmarks)","cited_arxiv_id":null,"evidence_quote":"Supplies the original FAN whose heatmaps the distilled FAN imitates and compares against."},{"cited_title":"Fsrnet: End-to-end learning face super-resolution with facial priors","cited_arxiv_id":null,"evidence_quote":"Defines FSRNet and FSRGAN, face SR baselines that use facial priors and are compared in the experiments."},{"cited_title":"Progressive growing of gans for improved quality, stability, and variation","cited_arxiv_id":null,"evidence_quote":"Motivates progressive growing of generator and discriminator, the training schedule adopted by the paper."},{"cited_title":"Improving the improved train- ing of wasserstein gans: A consistency term and its dual effect","cited_arxiv_id":null,"evidence_quote":"Delivers the WGAN gradient-penalty term that stabilizes the adversarial training."},{"cited_title":"Super-fan: Integrated facial landmark localization and super-resolution of real-world low resolution faces in arbitrary poses with gans","cited_arxiv_id":null,"evidence_quote":"Introduces the heatmap loss for aligning landmarks in super-resolved faces, which the paper adds to its objective."},{"cited_title":"Ultra-resolving face images by discriminative generative net- works","cited_arxiv_id":null,"evidence_quote":"Supplies URDGN, a GAN-based face super-resolution baseline used for quantitative and MOS comparison."}],"review_version":1}