{"id":"a3a2596f-d1f6-4a50-acc8-dc9c87302ef0","arxiv_id":"2502.09932","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"high","formal_verification":"none","parameter_count":3,"one_line_summary":"AffectSRNet combines landmark graph embeddings with an ESRGAN backbone to super-resolve faces while preserving facial expressions, and introduces the Emotion Consistency Metric (ECM) to evaluate emotion fidelity.","lead":"This paper proposes AffectSRNet, a face super-resolution network that uses facial landmark graphs and an emotion-preserving loss to upscale low-resolution faces while keeping expressions recognizable. It also introduces a new metric for measuring emotion preservation in super-resolved faces, and reports experiments on CelebA, FFHQ, and Helen.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The central emotion-fidelity claim is measured only by ECM, but ECM is an aggregate of marginal confidence histograms and mean entropy; it can stay unchanged when individual expressions are swapped, so the reported superiority may be an artifact of the metric.","rationale":"Working from the paper's own text, the strongest claim is the emotion-fidelity improvement. The architecture and ablations are coherent, and the qualitative comparisons are useful. The load-bearing assumption is not about consensus but about the validity of ECM. Since ECM aggregates marginal histograms and mean entropies, it is not a fidelity measure of individual expressions; it can be gamed by any transformation that preserves aggregate statistics. The paper explicitly says the emotion consistency losses are used as metrics and not for training, yet Eq. (2) includes Lhist; either Lhist is undefined here or the stated evaluation protocol is contradictory. A concrete validation against human labels or classifier label agreement would settle whether ECM tracks what it claims. Given the main claim depends on that unvalidated metric, the reader's REJECT with moderate confidence is appropriate; no change to the verdict is needed. I would note that the paper should not be judged by novelty alone; its central weakness is evidential, not stylistic.","tokens_in":16601,"tokens_out":4248,"duration_ms":56234,"concrete_test":"Perform a validation study on at least 100 images from CelebA and FFHQ at 4x and 8x: have human annotators label the emotion of HR originals and of SR outputs from AffectSRNet, DIC, and SPARNet, and compute a separate predicted-label agreement using DDAMFN++ on the same pairs. Then compute ECM for the same pairs. If AffectSRNet's ECM advantage is not accompanied by higher human or classifier label agreement with the HR originals, or if ECM does not correlate with label agreement across methods, the ECM is not a valid measure of emotion fidelity and the central claim is unsupported. As a quick internal check, permute DDAMFN++ confidence vectors per image before computing ECM; if ECM changes negligibly, label-level corruption is invisible to the metric.","verdict_should_be":"UNCHANGED","load_bearing_attack":"In Section III-A, ECM = αLH + log(Lconf). LH compares histograms of DDAMFN++ confidence scores per emotion class between original and super-resolved sets, and Lconf is the mean absolute difference of per-image entropy. Both are test-set-level marginal statistics. A model can pass both components while failing every individual emotion-preservation test: for example, if DDAMFN++ confidence scores of the super-resolved image are permuted across emotion classes per image, the class histograms and the mean entropy are essentially unchanged, yet the predicted emotion label is wrong on every image. The paper presents no calibration of ECM against human annotation, downstream FER accuracy, or predicted-label agreement. Since Tables I, II, and III report emotion fidelity exclusively through ECM, the central claim that AffectSRNet 'maintains the intensity and fidelity of facial expressions' rests on an unvalidated proxy. The inconsistency in Eq. (2), where Lhist appears in the training loss despite Section III-A stating emotion-consistency losses are metrics only, strengthens the concern that ECM is being optimized rather than independently measured. None of this attacks the architecture; it attacks the evidence for the headline claim.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes AffectSRNet, an emotion-aware face super-resolution network that augments an ESRGAN-style RRDB backbone with Mediapipe 3D facial landmarks processed by a GCN and fused through multimodal split-attention fusion blocks. The authors introduce an Emotion Consistency Metric (ECM) based on histogram differences and mean entropy differences of DDAMFN++ confidence scores between original and super-resolved test images, and report PSNR/SSIM/LPIPS/ECM comparisons on CelebA, FFHQ, and Helen at 4x and 8x upsampling. The central claims are that AffectSRNet outperforms existing FSR methods in emotion fidelity while remaining comparable in image quality, and that ECM is a useful new evaluation metric for emotion-aware super-resolution.","tokens_in":16887,"tokens_out":3916,"duration_ms":41440,"significance":"If the claims were established, the paper would address a real and under-explored problem: preserving facial expression information during face super-resolution for downstream FER applications. The architectural idea of injecting 3D landmark graph embeddings into a super-resolution backbone is plausible and the ablation study gives some evidence that the GCN+MSAF components help on standard image-quality metrics. However, the significance is undermined by the fact that the emotion-preservation claim rests entirely on an unvalidated metric, and by an internal inconsistency in which a component of that metric appears in the training loss. The manuscript does not include a user study, downstream FER accuracy evaluation, or predicted-label agreement analysis, so the headline claim is not currently supported by the evidence presented.","major_comments":[{"comment":"Section III-A states that the emotion consistency losses 'are used as metrics and not used during training,' yet Eq. (2) defines Ltotal = k1·Lh_pix + k2·Lhist + k3·G_phi_j + k4·LL2 with k2=20. The term Lhist is not defined in Section III-E, but the only histogram-based quantity introduced in the paper is LH from Section III-A, which is a component of ECM. If Lhist is indeed LH, then the evaluation metric is part of the training objective, making the reported ECM comparisons circular. This contradiction must be resolved: either remove Lhist from the loss and explain what Lhist is, or acknowledge that ECM is optimized during training and provide an independent evaluation.","section":"Section III-A and Eq. (2)"},{"comment":"The ECM is never validated as a measure of emotion preservation. It combines a test-set-level histogram difference and a mean absolute entropy difference, both computed from DDAMFN++ confidence scores. These marginal statistics can be unchanged even when per-image predicted emotion labels are wrong on every image, for example by permuting confidence scores across emotion classes within each image. The paper presents no calibration of ECM against human annotations, downstream FER accuracy, or per-image label agreement. Since the abstract and conclusion claim that AffectSRNet 'maintains the intensity and fidelity of facial expressions' and the experimental sections report emotion fidelity exclusively through ECM, the central claim is not established by the presented evidence.","section":"Section III-A and Tables I-III"},{"comment":"The text in Section IV-C.1 states that 'Across all three datasets, our method delivers results comparable to the state-of-the-art in PSNR, SSIM, and LPIPS while outperforming others in ECM.' This is contradicted by the FFHQ 4x row in Table I, where AffectSRNet reports LPIPS = 0.1260, while EDSR reports 0.0502 and FSRNet reports 0.0498. A roughly 2.5x worse LPIPS value is not 'comparable,' and it also places the method far behind two of the compared approaches. The claim of comparable visual quality is therefore not supported by the paper's own quantitative results.","section":"Table I and Section IV-C.1"},{"comment":"Several numerical results appear inconsistent or unexplained, which weakens confidence in the reported comparisons. For example, in Table I the Helen LPIPS for SRCNN is 0.556, which is dramatically worse than bicubic interpolation (0.1771) — an unusual outcome for a trained SR model. In the same table, SPARNet on FFHQ 4x shows LPIPS = 0.1878, far worse than EDSR and FSRNet, yet the qualitative discussion describes SPARNet as generally retaining texture. The paper should provide implementation details, error bars or statistical significance tests, and an explanation for these outliers before the quantitative comparisons can be taken at face value.","section":"Section IV-C.1 and Tables I-II"}],"minor_comments":[{"comment":"The contribution bullet 'We provide comprehensive quantitative as well as quantitative comparisons' uses 'quantitative' twice; the second occurrence should presumably be 'qualitative.'","section":"Section I-A"},{"comment":"The perceptual loss is written only as the Gram matrix element G_phi_j(x)_c,c'; the actual style reconstruction loss should be a sum over channel pairs, but this sum is never defined. Please clarify the exact loss term used in Eq. (2).","section":"Section III-E.2"},{"comment":"The L2 node-embedding loss in Eq. (1) refers to graphs G1 and G2, but neither graph is defined in the text. Presumably these are the landmark graphs extracted from the low-resolution input and the high-resolution target, but this must be stated explicitly.","section":"Section III-E.3 and Eq. (1)"},{"comment":"The ablation table does not state the upsampling factor, although the text says the ablation is performed at 8x. Please add the scale to the table caption or header.","section":"Table III"},{"comment":"There are apparent reference errors: the same FSRNet paper appears as both [7] and [8], and the pose-estimation reference [24] is later cited for a face attention unit in Section II-C.1, which appears to be a different work. Please correct the citation numbering and deduplicate entries.","section":"References"},{"comment":"The caption says 'SRCN' instead of 'SRCNN.'","section":"Fig. 4 caption"}],"recommendation":"reject","confidential_remarks":"The central problem is that the paper's main contribution — emotion-preserving super-resolution — is evaluated with a metric that is both unvalidated and, according to Eq. (2), likely part of the training objective. The FFHQ 4x LPIPS result also directly contradicts the 'comparable visual quality' claim. These are load-bearing issues rather than presentation fixes. The architecture may have merit, but the evidence presented does not support the headline claim, and repairing this would require a substantial new evaluation study rather than a minor revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The core idea here is reasonable: inject 3D landmark structure into an ESRGAN backbone via a GCN and MSAF fusion to keep expressions intact. That integration is incremental over landmark-guided FSR, but it is a fair thing to try, and the attempt to define an emotion consistency metric (ECM) is useful in principle. I'll give credit for that.\n\nThe problem is the evidence. ECM is never calibrated against human perception, downstream FER accuracy, or predicted-label agreement. It is a test-set-level aggregate of confidence histograms and mean entropy, so it can stay flat while individual expressions are swapped. The stress-test note makes that point concretely, and it lands. If a model messes up every expression but preserves the marginal histograms, ECM would not care. That means the reported superiority in Tables I and II may be an artifact of the metric.\n\nWorse, there is an internal contradiction. Section III-A states plainly that the emotion consistency losses are metrics only, not used in training. But Eq. (2) includes Lhist with weight 20, and Lhist is never defined anywhere. The only histogram-like loss in the paper is the LH inside ECM. If that is what Lhist refers to, then the evaluation metric is part of the training objective, and the main empirical claim is circular. That is not a typo you can wave away; it is load-bearing.\n\nImage quality results are also mixed. On FFHQ 4x, their LPIPS is 0.126, worse than EDSR and FSRNet at around 0.050. The abstract promises \"comparable\" visual quality; the tables do not fully support that. There is also duplicated text between the two related-work subsections, and no code is released beyond \"upon a reasonable request,\" which is not reproducible.\n\nWho is this for? Someone working on face SR or FER might take the architecture idea and run with it, but not the evaluation protocol as presented. I would not cite this in its current form.\n\nFor peer review: I would send it to referees rather than desk reject, because the questions it raises about metric validation are important and the architecture is worth examining. But I would expect heavy revision: retrain without any emotion loss in the objective, validate ECM against real labels or human raters, and fix the Lhist contradiction. As is, the headline claim does not hold.","headline":"Sensible architecture, but the emotion-preservation claim rests on an unvalidated metric and a training-loss contradiction that undermines the main result.","tokens_in":17377,"tokens_out":2270,"would_cite":false,"duration_ms":26086,"reading_group":"maybe","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper introduces AffectSRNet, a face super-resolution network that uses graph embeddings of facial landmarks to upscale low-resolution faces while preserving the intensity and fidelity of their expressions, and claims it outperforms…","keywords":["face super-resolution","facial expression recognition","emotion preservation","graph convolutional network","multimodal fusion","emotion consistency metric","low-resolution faces","landmark priors"],"falsifier":"Train or fine-tune AffectSRNet with the histogram-loss term Lhist removed (set k2=0 in Eq. (2)) and compare ECM against the published numbers; if the emotion-fidelity gain largely disappears, the reported advantage comes from optimizing the metric rather than from the landmark-graph architecture. Alternatively, replace DDAMFN++ with a different, independently trained FER model or with human emotion ratings and check whether AffectSRNet still ranks first on the resulting ECM.","tokens_in":16429,"feed_emoji":"😀","tokens_out":4875,"duration_ms":45225,"temperature":0.7,"pith_summary":"The paper tries to establish that a face super-resolution network can restore high-resolution facial detail from low-resolution input without washing out the emotion on the face. It argues that facial landmark geometry, encoded as graph embeddings and fused into a standard super-resolution backbone, is the right prior for keeping expressions intact. To back this up, it introduces a new Emotion Consistency Metric that compares the confidence distributions of a pretrained expression classifier on original and super-resolved faces. The authors report that AffectSRNet beats six existing super-resolution methods on this metric on CelebA, FFHQ, and Helen, while staying comparable on PSNR, SSIM, and LPIPS.","feed_headline":"AffectSRNet beats face-SR baselines at preserving emotion","feed_subtitle":"Graph embeddings of facial landmarks plus a new emotion consistency metric drive the gain.","key_machinery":"The central machinery is a graph-embedding pathway: 478 facial landmarks extracted by Mediapipe are connected by hand-defined edges linking eyes, lips, and cheeks, passed through stacked Graph Convolutional Network layers, and fused into the RRDB super-resolution backbone at three points via Multimodal Split Attention Fusion (MSAF) blocks. The other load-bearing piece is the Emotion Consistency Metric (ECM), defined as ECM = 0.5*LH + log(Lconf), where LH is the histogram loss between class-confidence histograms of original and super-resolved images and Lconf is the mean absolute difference in predictive logit entropy; this metric is meant to quantify how much emotional content survives upscaling.","core_discovery":"On the paper's own terms, the discovery is that expression-preserving super-resolution can be achieved by injecting facial-landmark graph embeddings into an RRDB-based super-resolution network through a multimodal split-attention fusion block. The key result is that this architecture lowers the Emotion Consistency Metric on all three benchmark datasets at both 4x and 8x upscaling compared with SRCNN, EDSR, FSRNet, DIC, and SPARNet, with no consistent loss in standard image-quality metrics. The paper further claims that the new ECM, computed as a log-weighted sum of histogram loss and average predictive-entropy difference from the DDAMFN++ classifier, is a useful tool for evaluating emotion preservation in super-resolved faces.","pith_inferences":["A direct check the authors do not report: whether ECM improvements translate to higher downstream facial-expression recognition accuracy on an independent benchmark, or only to closer agreement with DDAMFN++.","The paper says emotion-consistency losses are used only as metrics, yet Eq. (2) explicitly includes Lhist in the training objective; if the histogram loss was active during training, the reported ECM wins could reflect metric overfitting rather than genuine expression preservation.","The same graph-embedding prior could plausibly be adapted to preserve other facial attributes such as age, gaze, or identity during super-resolution, extending the core idea beyond emotion."],"forward_implications":["If AffectSRNet's emotion-fidelity gain holds, low-resolution surveillance footage could be upscaled before expression analysis without sacrificing FER accuracy.","The extensible loss function could be fine-tuned into existing face super-resolution networks, giving them an emotion-preserving mode.","The Emotion Consistency Metric gives future work a quantitative target for emotion-aware super-resolution, beyond PSNR and SSIM.","At 8x upscaling, where generic methods blur away expression cues, the graph-embedding prior is claimed to keep fine mouth and eye structure intact."],"supporting_citations":[{"why":"Supplies the RRDB-based super-resolution backbone that AffectSRNet modifies.","marker":"[82]"},{"why":"Provides the 478 facial landmarks used as the structural prior.","marker":"[33]"},{"why":"Defines the graph convolutional layers that turn landmarks into graph embeddings.","marker":"[38]"},{"why":"Defines the Multimodal Split Attention Fusion block that merges graph embeddings into the SR backbone.","marker":"[69]"},{"why":"The DDAMFN++ model whose confidence outputs are used to compute the Emotion Consistency Metric.","marker":"[96]"},{"why":"CelebA, one of the three benchmark datasets used for training and evaluation.","marker":"[55]"},{"why":"FFHQ, one of the three benchmark datasets used for training and evaluation.","marker":"[32]"},{"why":"Helen, one of the three benchmark datasets used for training and evaluation.","marker":"[43]"}],"fun_headline_variants":["AffectSRNet preserves emotion in low-res face upscaling","Landmark graph embeddings keep expressions intact in face SR","New emotion consistency metric validates AffectSRNet's gains","Emotion-aware SR: graph fusion beats standard face SR"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The entire claim rests on the assumption that the Emotion Consistency Metric, which depends on the confidence outputs of one pretrained expression classifier (DDAMFN++), genuinely tracks whether a human would still read the same emotion in the super-resolved face.","fun_headline_variants_meta":{"raw":{"variants":["AffectSRNet preserves emotion in low-res face upscaling","Landmark graph embeddings keep expressions intact in face SR","New emotion consistency metric validates AffectSRNet's gains","Emotion-aware SR: graph fusion beats standard face SR"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000239,"raw_usage":{"total_tokens":1516,"prompt_tokens":947,"completion_tokens":569,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":563,"completion_tokens_details":{"reasoning_tokens":502}},"tokens_in":563,"tokens_out":569,"duration_ms":6413,"temperature":1.0,"reasoning_tokens":502,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T20:00:11.143873+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train or fine-tune AffectSRNet with the histogram-loss term Lhist removed (set k2=0 in Eq. (2)) and compare ECM against the published numbers; if the emotion-fidelity gain largely disappears, the reported advantage comes from optimizing the metric rather than from the landmark-graph architecture. Alternatively, replace DDAMFN++ with a different, independently trained FER model or with human emotion ratings and check whether AffectSRNet still ranks first on the resulting ECM.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the RRDB-based super-resolution backbone that AffectSRNet modifies."},{"cited_title":"Kartynnik, A","cited_arxiv_id":null,"evidence_quote":"Provides the 478 facial landmarks used as the structural prior."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines the Multimodal Split Attention Fusion block that merges graph embeddings into the SR backbone."},{"cited_title":"Zhang, Y","cited_arxiv_id":null,"evidence_quote":"The DDAMFN++ model whose confidence outputs are used to compute the Emotion Consistency Metric."},{"cited_title":"Karras, S","cited_arxiv_id":null,"evidence_quote":"FFHQ, one of the three benchmark datasets used for training and evaluation."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Helen, one of the three benchmark datasets used for training and evaluation."}],"review_version":1}