{"id":"08029279-3234-4519-8abe-f699bbb944df","arxiv_id":"2411.19246","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"Face2QR generates scannable, aesthetic QR codes that preserve face identity by integrating face and QR control in a Stable Diffusion pipeline, reshuffling QR modules, and optimizing in latent space.","lead":"A new three-stage pipeline generates QR codes that contain a person's face while keeping the code scannable by phones. The method combines face-preserving image generation with QR module reshuffling and scan-robustness optimization.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The IDRS stage assumes, without proof or failure-rate data, that freezing face/marker modules still leaves some assignment of remaining QR modules that decodes to the target message; if this feasibility premise fails, the pipeline's general scannability and SOTA claims lose their foundation.","rationale":"The paper's strongest claim is that Face2QR produces ID-preserved, scannable QR codes at SOTA level. The mechanism that supposedly reconciles face identity with QR data is IDRS, and Eqs. 3-4 are the formal statement of that mechanism. The paper treats the constrained optimization as always solvable: it says 'to ensure' the result decodes, it regenerates the ECC, but gives no feasibility argument, no ECC level, and no failure statistics. For QR codes, the placement of data and ECC codewords is fixed; freezing face modules fixes some of those bits. If the forced bits conflict with the target message beyond the ECC correction capability, no remaining-module assignment can fix the code. This is not merely a missing theoretical nicety: the paper's own Figure 2 shows the stage-2 result still has 17.55% error, so the blueprint is not guaranteed to be lossless, and the final scanning success is demonstrated on only 20 selected images. A second concern, the 0.01 gap in d and the absence of error bars, would weaken the SOTA comparison, but it is secondary because even a perfect evaluation of the reported 20 images would not establish that the pipeline works generally. The proposed concrete test directly settles the feasibility question by searching for a valid module assignment under the exact constraints, over a larger random sample. I agree with the reader's identification of this as the weakest assumption, and the reader's CONDITIONAL verdict is appropriate; this stress-test does not move it.","tokens_in":17164,"tokens_out":6734,"duration_ms":65597,"concrete_test":"Reproduce the IDRS step exactly on the paper's protocol (SDXL/InstantID/QR Controller, version 5, 1024x1024) for the 20 reported test images plus 100 additional random face/prompt/message triples. For each input, compute the face mask, binarize the stage-1 output, and solve the constrained assignment of Eqs. 3-4 by exhaustive search over all 8 QR masks and all four ECC levels (L/M/Q/H), using a standard QR library to test whether any non-face, non-marker module assignment makes the blueprint decode to the target message. If the fraction of feasible solutions is below 100% or requires an ECC level different from the one used by the pipeline, the general scannability claim must be qualified and the failure rate reported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim rests on the IDRS feasibility premise in Eqs. 3-4. For a version-5 QR code, data and Reed-Solomon ECC codewords occupy fixed module positions determined by the QR specification. Freezing the face region and markers fixes the brightness of those modules, which fixes some data-codeword bits. If those frozen bits disagree with the target message in more data codewords than the ECC can correct, no rearrangement of the remaining modules and no 'regenerated' ECC can make the blueprint decode to the target message. The paper never states the ECC level, never analyzes this constraint, never reports an IDRS failure rate, and never shows that the blueprint itself decodes; it only reports average error rates after later stages. Because the scanning test uses only 20 selected images and the objective table shows a 0.01 difference on d, the reported success heavily depends on this unverified step. If IDRS frequently has no feasible solution, the method only works on favorable inputs and the claimed general SOTA performance is not established.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Face2QR, a three-stage, training-free pipeline for generating aesthetic QR codes that preserve a person's face identity while remaining scannable. Stage 1 (IDQR) uses SDXL with InstantID and a QR Controller to synthesize an initial image. Stage 2 (IDRS) binarizes this image, freezes the modules in the face and marker regions, reshuffles the remaining modules, and regenerates the error-correction codewords to produce a scannable blueprint that guides a second diffusion pass. Stage 3 (IDSE) optimizes the latent code with a spatially adaptive code loss and an aesthetic content loss to enhance scanning robustness while limiting modifications to the face region. Experiments report scan success rates above 94% across three decoders and three display sizes for 20 generated codes, a user study favoring Face2QR, and objective comparisons using ArcFace-based identity distance and AesBench aesthetic scores, leading to a claimed state-of-the-art performance.","tokens_in":17336,"tokens_out":3947,"duration_ms":34953,"significance":"If the central claims hold, Face2QR is a practical contribution: it is a single pipeline that combines face identity, arbitrary aesthetic backgrounds, and scannability, and the paper includes real-world scanning tests with multiple phone decoders, a user study, ablations of the main components, and a code-release URL. The design of the IDRS blueprint and the spatially dynamic loss function are reasonable and clearly motivated by the failure modes of prior work. However, the evidence as presented has several load-bearing gaps—the feasibility of the IDRS constrained optimization is not established, the scanning evaluation uses only 20 hand-picked codes without statistical significance, and the objective error metric may partially mirror the optimization objective—so the SOTA claim is not yet fully supported.","major_comments":[{"comment":"The feasibility of the IDRS constrained optimization is an unverified premise that the whole pipeline depends on. The constraints freeze the module values in the face and marker regions, which fixes the bits of some data and Reed–Solomon codewords at their physical positions; if those fixed codewords already contain more errors than the ECC can correct, no rearrangement of the remaining modules and no 'regenerated' error correction can make the blueprint decode to the target message. The paper never states the QR ECC level, never reports the fraction of inputs for which a feasible rearranged blueprint exists, and never reports that the blueprint itself decodes before the later stages. Because the scannability claim relies on this step, the authors should add a feasibility analysis over a substantial random sample of face images and messages, report the IDRS failure rate, and verify that the blueprint decodes with a standard decoder.","section":"§3.1, Eqs. (3)–(4) and §4.1"},{"comment":"The scannability evidence is based on only 20 hand-picked codes and 50 trials, with no error bars, confidence intervals, or statement of how the codes were selected. The NeurIPS checklist confirms that no statistical significance is reported. As the paper's headline 'average success rate exceeding 94%' and its comparison with Text2QR rely entirely on this table, the authors should report per-code variance, use a larger and more clearly sampled set, and ideally include a significance test against the baseline.","section":"§4.3, Table 5"},{"comment":"The metric d is defined in §4.1 as 'face feature distance' but is computed as a cosine similarity, for which higher values mean better identity preservation; the name is misleading. Moreover, the 0.01 advantage over ArtCoder (0.51 vs. 0.50) is within likely noise, especially because the reported averages are over only 100 samples without error bars. Please clarify the direction and units of d, report standard deviations or confidence intervals, and avoid claiming SOTA on this metric without a significance test.","section":"§4.3, Table 6"},{"comment":"The objective error metric e appears to be computed with the same simulated Gaussian-decoder model that drives the adaptive code loss Lc. If e is computed this way, the quantitative scannability evaluation partly mirrors the optimization objective and may overstate how well the output decodes with real scanners. The authors should specify exactly how e is obtained, and if the simulated decoder is used, add an independent evaluation using the same real decoders as in Table 5 to confirm that the reported error rates transfer to practice.","section":"§3.2.2, Eqs. (9)–(11) and §4.1"}],"minor_comments":[{"comment":"The module is called 'IDRS' in most of the paper but 'IDRE' in these sections; please unify the name throughout.","section":"§3.1 and §4.4"},{"comment":"The heading 'Quantitative Comparsion' should be 'Quantitative Comparison'; also fix the typos 'Scannibility', 'Harmonziation', and 'inherit conflict' elsewhere in the text.","section":"§4.3 heading"},{"comment":"The caption notes '[Key: Best]' but the table does not use boldface or another marker to show which entry is best, which makes the table hard to read.","section":"Table 6"},{"comment":"The definition of e says error modules are counted 'excluding finder and alignment pattern areas,' but the ablation tables also report ef within the face region; please clarify whether e includes modules in the face region or excludes them, since the two statements can be read as contradictory.","section":"§4.1"},{"comment":"Reference [8] is cited as 'Huang et al.' with no initial and an incomplete author list; please provide the full citation for the AesBench tool.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The paper is a plausible and well-motivated application contribution, and the real-world scanning tests are a genuine strength. However, the central SOTA claim currently rests on unverified feasibility of the IDRS step, a small hand-picked scanning set with no significance testing, and an objective metric that may be circularly tied to the optimization loss. These are fixable in revision, so I would not reject, but I would not accept the current evidence. Please also ensure the code release actually happens and that the ECC level and feasibility analysis are included, as the checklist alone is not sufficient verification."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a training-free pipeline that combines InstantID with the QR Controller from Text2QR, plus one genuinely new idea—IDRS, which freezes the QR modules under the face and the finder/alignment markers, then reshuffles the remaining modules and regenerates the Reed-Solomon error correction so the blueprint still decodes. That is a real new technique, and the paper earns credit for testing with actual phone scans on a monitor rather than only simulated decoders. The 94%+ success rates across three decoders and three display sizes are the strongest evidence in the paper.\n\nThe soft spots are mostly in the quantitative claims. The headline SOTA rests on face feature \"distance\" d of 0.51 versus 0.50 for ArtCoder and 0.43 for Text2QR, with no error bars or significance tests; the checklist explicitly says No. Since d is actually a cosine similarity—the paper calls it a distance but higher is better—a 0.01 difference is meaningless without variance. The scanning test uses 20 hand-picked codes: enough to show the pipeline works, not enough to support a general SOTA claim.\n\nThe deeper concern is the one the stress-test note flags. IDRS assumes that after freezing face and marker modules, the remaining modules can be rearranged with regenerated ECC to decode the target message. For a version-5 QR code, that is a constrained feasibility problem. The paper states no ECC level, reports no IDRS failure rate, and never shows the blueprint itself decoding—only later-stage error rates. If IDRS frequently yields no feasible assignment, the method only works on favorable inputs. That said, the reported intermediate error rates dropping from 43.85% to 17.55% to 0.01% suggest IDRS succeeds on at least the shown examples, but those are anecdotes.\n\nMinor but real: the abstract promises code at a URL while the checklist says release comes after a patent process; VGG-19 is described as pre-trained on MS-COCO, which is likely wrong; and the IDSE error metric is computed with the same simulated Gaussian decoder that drives the training loss, so it is partly circular. All addressable.\n\nBottom line: this is a serious applied paper, honestly presented with a limitations section and bad cases. The IDRS idea deserves referee time. I would accept it for peer review and ask for a feasibility analysis, failure rate, and statistical rigor. I'd bring it to a reading group and would cite it if I worked on QR aesthetics.","headline":"Solid engineering paper with a genuinely new IDRS trick, but the SOTA claim and the IDRS feasibility premise need more support than the paper provides.","tokens_in":17956,"tokens_out":2645,"would_cite":true,"duration_ms":23041,"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":"Face2QR claims that a training-free three-stage pipeline can generate aesthetically styled QR codes that preserve a person's facial identity and scan reliably, with an average face feature distance of 0.51 and scanning success above 94%.","keywords":["aesthetic QR codes","face identity preservation","scannability","Stable Diffusion","control networks","QR module reshuffle","latent code optimization","Reed-Solomon error correction"],"falsifier":"Run the IDRS stage alone on a set of face images and QR payloads with varying lengths, and count how often the constrained shuffle, with the face and marker modules frozen, produces a decodable blueprint before any IDSE refinement. If any realistic face mask leaves too few free modules for the message's error-correction requirement, so that the stage silently raises the error-correction level or must alter the face region, the universality of the pipeline is falsified. Independently, a neutral reproduction of the reported 0.51 average face feature distance using the paper's stated settings would confirm or refute the identity-preservation claim.","tokens_in":16896,"feed_emoji":"🎨","tokens_out":11159,"duration_ms":84487,"temperature":0.7,"pith_summary":"The paper tries to establish that a personalized QR code does not have to trade face identity, visual appeal, and scannability against one another. It proposes Face2QR, a three-stage pipeline that generates an initial aesthetic image from a face photo, a text prompt, and a QR message; rebuilds the QR blueprint with the face region frozen; and then refines the latent image with a region-aware loss so the code is reliably decodable. If the claim is right, the practical pay-off is that anyone can produce a decorative, face-recognizable QR image that a phone camera actually scans, with no per-user training. The paper reports an average face feature distance of 0.51, an aesthetic score of 90.1 on AesBench, and scanning success rates above 94% across three decoders and three display sizes. The importance of the claim is that it turns scannability from a strict global constraint into a repairable property of a small set of QR modules.","feed_headline":"Face2QR keeps face IDs in QR codes that scan 94% of the time","feed_subtitle":"Three training-free stages protect the face, rebuild the QR pattern, and sharpen scan reliability.","key_machinery":"The central mechanism is the split of QR modules into protected and malleable sets. IDQR (ID-refined QR integration) uses a Stable-Diffusion-based generator steered by an identity-control network and a QR Controller. IDRS (ID-aware QR ReShuffle) binarizes the generated image, freezes the modules that fall inside the face mask and those belonging to the finder/alignment markers, rearranges the remaining modules so the code matches the target message, and regenerates the Reed-Solomon error correction for the rearranged layout. IDSE (ID-preserved Scannability Enhancement) strengthens the marker patterns and then optimizes the latent code with an adaptive code loss, which applies Gaussian kernels of different widths and different weights to face-region modules versus background modules, together with an aesthetic content loss based on the $L^2$-Wasserstein distance between VGG-19 feature maps of the output and the pre-enhancement image. The entire pipeline is training-free, so the same three stages can be reused with different base generators or identity encoders.","core_discovery":"On its own terms, the paper's discovery is that the three-way conflict between face identity, aesthetics, and QR scannability can be resolved by separating the control signals across three stages instead of asking one model to satisfy all constraints at once. Stage one, IDQR, blends the face into a styled background under two control networks, one for identity and one for the QR pattern. Stage two, IDRS, fixes the QR modules under the face mask and under the finder/alignment markers, rearranges the remaining modules to agree with the target message, and regenerates the error-correction codewords so the new blueprint still decodes. Stage three, IDSE, strengthens the marker patterns and optimizes the latent code with a spatially dynamic loss that is gentle on the face and strict on the background. The authors argue that this division removes the main failure mode of earlier aesthetic-QR methods, which either distorted the face, left the code unscannable, or both.","pith_inferences":["A natural extension is to protect any region that must remain visually clean, such as a logo, signature, or product image, by freezing its QR modules and reshuffling only the rest, provided the protected region leaves enough free modules for the required error correction.","The paper leaves the feasibility of the IDRS rearrangement unquantified; one could measure, across many face masks and payloads, how often the frozen regions force the regenerated error correction to fail or force a change in the QR error-correction level.","If such codes become common, they double as visible biometric tokens, so the same technical success raises a privacy question the paper does not address: the QR image itself now carries a recognizable face in plain sight."],"forward_implications":["If the reported results hold, a single pass through the three stages produces QR images that common phone decoders read successfully in more than 94% of trials at display sizes from 3 cm to 7 cm.","Because the pipeline is training-free, new base generators or better identity encoders could be swapped in without retraining the three-stage structure itself.","The stage-by-stage error reduction, from 43.85% after stage one to 0.01% after stage three, implies that scannability can be treated as a post-generation repair problem rather than a single end-to-end constraint.","The adaptive loss result implies that protecting a semantically important region during latent optimization does not force a global quality-versus-scannability trade-off: different regions can be optimized with different strictness."],"supporting_citations":[{"why":"InstantID supplies the identity-preserving control network that keeps the generated face close to the input photo.","marker":"[40]"},{"why":"Text2QR provides the QR Controller blueprint guidance and the simulated-decoder loss design that IDSE adapts.","marker":"[43]"},{"why":"ArcFace is the recognition model used to compute the face feature distance d, the paper's main identity metric.","marker":"[7]"},{"why":"AesBench supplies the aesthetic scoring tool Ba used for the objective quality comparison.","marker":"[8]"},{"why":"The cited face-analysis model produces the face mask Mf that defines which QR modules are protected in IDRS and IDSE.","marker":"[14]"},{"why":"Reed-Solomon error correction is the encoding mechanism regenerated during the QR module reshuffle so the modified blueprint still decodes to the message.","marker":"[29]"},{"why":"VGG-19 provides the feature layers for the aesthetic content loss in IDSE.","marker":"[34]"},{"why":"VQ-VAE is the latent autoencoder whose encoder and decoder are used for the latent code optimization in IDSE.","marker":"[39]"}],"fun_headline_variants":["Face2QR: QR codes that keep your face and scan","Face2QR: Three-stage pipeline for face-preserving scannable QR codes","Face2QR: Aesthetic QR codes that preserve identity and scannability","Face2QR: Solving the face-QR tradeoff with separated controls","Face2QR: QR codes that look like you and scan every time"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that, after keeping every black-and-white square of the QR code that overlaps the face or the three corner markers exactly as it is, the remaining squares can always be rearranged, with new error-correction data, to form a valid QR code for the intended message; the paper gives no feasibility analysis, no statement of how much error correction is used, and no failure rate for this step.","fun_headline_variants_meta":{"raw":{"variants":["Face2QR: QR codes that keep your face and scan","Face2QR: Three-stage pipeline for face-preserving scannable QR codes","Face2QR: Aesthetic QR codes that preserve identity and scannability","Face2QR: Solving the face-QR tradeoff with separated controls","Face2QR: QR codes that look like you and scan every time"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000586,"raw_usage":{"total_tokens":2765,"prompt_tokens":968,"completion_tokens":1797,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":584,"completion_tokens_details":{"reasoning_tokens":1700}},"tokens_in":584,"tokens_out":1797,"duration_ms":13321,"temperature":1.0,"reasoning_tokens":1700,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T10:24:21.896227+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the IDRS stage alone on a set of face images and QR payloads with varying lengths, and count how often the constrained shuffle, with the face and marker modules frozen, produces a decodable blueprint before any IDSE refinement. If any realistic face mask leaves too few free modules for the message's error-correction requirement, so that the stage silently raises the error-correction level or must alter the face region, the universality of the pipeline is falsified. Independently, a neutral reproduction of the reported 0.51 average face feature distance using the paper's stated settings would confirm or refute the identity-preservation claim.","supporting_citations":[{"cited_title":"Text2qr: Harmonizing aesthetic customization and scanning robustness for text-guided qr code generation","cited_arxiv_id":null,"evidence_quote":"Text2QR provides the QR Controller blueprint guidance and the simulated-decoder loss design that IDSE adapts."},{"cited_title":"Insightface: A 2d and 3d face analysis project, 2023","cited_arxiv_id":null,"evidence_quote":"The cited face-analysis model produces the face mask Mf that defines which QR modules are protected in IDRS and IDSE."},{"cited_title":"Polynomial codes over certain finite fields","cited_arxiv_id":null,"evidence_quote":"Reed-Solomon error correction is the encoding mechanism regenerated during the QR module reshuffle so the modified blueprint still decodes to the message."},{"cited_title":"Very deep convolutional networks for large-scale image recogni- tion","cited_arxiv_id":null,"evidence_quote":"VGG-19 provides the feature layers for the aesthetic content loss in IDSE."}],"review_version":1}