{"id":"4b7b8f2a-5fcc-4eab-91bc-d284fd827481","arxiv_id":"2411.16446","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":8,"one_line_summary":"VQ-SGen encodes each stroke as discrete shape and location codes and generates them autoregressively, producing more coherent and diverse creative sketches than prior methods.","lead":"VQ-SGen is a two-stage sketch generation method that learns compact vector-quantized codes for individual strokes, then generates coherent sketches stroke by stroke with an autoregressive Transformer. It reports better fidelity and diversity than prior sketch generators on the CreativeSketch benchmark and supports text, class label, and sketch completion conditioning.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Metric validity: FID/GD/CS in a QuickDraw-Inception space do not establish creative superiority; a memorization baseline could match Table 1, and the only human study compares against a single baseline.","rationale":"The central claim is empirical: VQ-SGen surpasses SOTA for creative sketch generation. The most load-bearing premises are (1) the QuickDraw-Inception metrics measure the quality/creativity targeted by the task, and (2) the baselines are evaluated fairly. Premise (1) is least secure. FID/GD/CS are distributional statistics in a feature space trained on standard QuickDraw doodles; they reward closeness to the CreativeSketch training distribution and diversity in that space. A memorizing system would score very well on these statistics, so they cannot by themselves establish generation of novel creative sketches. The paper's own discussion of the CB CS result (higher CS indicates \"simplicity\") implicitly concedes that the metrics are not monotone in the desired property. The user study is the only direct evidence of creativity, but it covers only DoodleFormer among the compared baselines, with no significance testing. I therefore do not see an internal inconsistency in the method; the soft spot is the validity of the evaluation for the stated claim. The memorization-control experiment would settle whether the Table 1 margins are meaningful; if the control fails, the superiority claim needs to be reargued with human evaluation against all baselines or a creativity-specific metric. This does not change the reader's conditional verdict: the paper is a plausible advance, but the claimed superiority is not yet fully established.","tokens_in":16075,"tokens_out":11462,"duration_ms":110827,"concrete_test":"Run a memorization control: draw the same number of sketches as the evaluation set by random resampling from the CreativeSketch training set (or, to be stricter, by retrieving the nearest training sketch to each generated sketch), and compute FID/GD/CS/SDS with the same QuickDraw Inception features. If this copy baseline matches or beats VQ-SGen's Table 1 numbers, the quantitative superiority claim is invalidated as evidence of creative generation. If it underperforms clearly, the metric-based claim is at least not trivially circular.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that VQ-SGen \"surpasses existing state-of-the-art techniques\" on the CreativeSketch dataset. The only quantitative support is Table 1, computed in the feature space of a QuickDraw3.8M-trained Inception network. These metrics measure distributional fidelity and diversity in that fixed feature space, not creativity, novelty, or aesthetic quality. FID in particular is minimized by reproducing the training distribution; a baseline that returns random training sketches would likely score very well on FID/GD/CS while generating nothing new. The paper's own handling of the Creative Birds CS scores (Ours 0.53 vs DoodlerGAN 0.69 and DoodleFormer 0.57) concedes that a higher CS may indicate only \"simplicity\" rather than quality, which undermines the assumption that the Table 1 metrics are monotonically aligned with the claimed property. The only direct human evaluation (Sec. 4.3) compares against DoodleFormer and real data, not against DoodlerGAN or SketchKnitter, and reports no statistical significance. Thus the evidence that VQ-SGen \"surpasses\" SOTA in the sense claimed is not yet settled: the quantitative margin may reflect metric artifacts rather than creative generation.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes VQ-SGen, a two-stage generative model for creative sketch drawing. In stage one, each stroke is represented as a triplet of shape, location, and label; the shape and location are separately vector-quantized through a VQ-VAE-style codebook. In stage two, a cascaded autoregressive Transformer generates stroke labels and then shape/location codes, enabling unconditional, class-conditioned, text-conditioned, and completion-based sketch generation. Experiments on CreativeSketch (Creative Birds and Creative Creatures) report improved FID, GD, CS, and SDS over DoodlerGAN, DoodleFormer, and SketchKnitter, together with ablation studies and a user study.","tokens_in":16355,"tokens_out":3819,"duration_ms":39139,"significance":"If the empirical claims hold, VQ-SGen is a useful step for creative sketch generation: it provides a compact discrete stroke representation that captures shape while decoupling position, and it demonstrates that such a representation can support a simple autoregressive generator with several conditioning modes. The paper has several strengths: the two-stage design is clearly described, the ablations in Table 2 directly test the contribution of the VQ representation, decoupling, and label conditioning, and the code-space visualization in Fig. 8 shows semantically meaningful clusters that justify the tokenized formulation. The main limitation is that the central superiority claim rests on point estimates of FID/GD/CS/SDS and on a user study that is partial and unaccompanied by significance testing; the metric validity concern raised below is therefore load-bearing rather than cosmetic.","major_comments":[{"comment":"All quantitative results are reported as single point estimates. The abstract and §4.1 claim that the method 'significantly outperforms' prior work, but without multiple seeds, standard deviations, confidence intervals, or significance tests, the reported margins (FID 15.78 vs 17.48 on Creative Birds and 17.61 vs 20.43 on Creative Creatures) cannot be distinguished from run-to-run variation. Please report mean±std over at least three independent training runs and apply a paired test between the methods.","section":"§4.1, Table 1"},{"comment":"The FID, GD, CS, and SDS metrics are all computed in the feature space of an Inception network trained on QuickDraw3.8M. These metrics measure distributional fidelity and diversity in that fixed feature space, not creativity or aesthetic novelty; a baseline that simply returned random training sketches would likely score very well on FID/GD/CS while generating nothing new. The paper itself concedes in §4.1 that a higher CS may reflect 'simplicity of the generation' rather than quality, which undercuts the assumption that the Table 1 metrics are monotonically aligned with the claimed property. Please add a memorization/retrieval baseline, and either include a second metric family (e.g., in a different feature space) or temper the 'surpasses state-of-the-art' claim to 'improves on these metrics in this feature space'.","section":"§4, 'Evaluation metrics'"},{"comment":"The user study compares VQ-SGen only against DoodleFormer and against real dataset sketches, not against DoodlerGAN or SketchKnitter, and no confidence intervals or significance tests are reported. Since the paper concludes from this study that the method 'consistently performs favorably against SoTA methods', the evidence is incomplete. Include all three baselines in the user study, or restrict the claim to the DoodleFormer comparison; in either case, report participant-level statistics.","section":"§4.3, Fig. 6"},{"comment":"The statement that 'all three methods are trained using their default parameters on our dataset' is underspecified. If the baseline implementations were used with their released defaults, those defaults may have been tuned for other datasets, which could systematically disadvantage the baselines. Please clarify whether any hyperparameter validation was performed for the baselines and, if not, state this as a limitation of the comparison.","section":"§4.1, 'Comparison'"}],"minor_comments":[{"comment":"The dimensions of z_i^s and c_j are not given in the main text; please specify them and state the distance metric used for the nearest-code assignment.","section":"§3.1, Eq. (2)"},{"comment":"The balancing weight α multiplies both the codebook and commitment terms; the design choice and the value α=0.8 are reported only in the supplementary. Please state the loss weighting in the main text or move the value to the main description.","section":"§3.1, Eq. (3)"},{"comment":"The phrase 'build a space for the shape location' reads as a typo; presumably 'shape and location' is intended.","section":"§3.1"},{"comment":"References [1] and [2] are the same SketchINR paper, and [14] and [15] are the same paper; these duplicates should be consolidated.","section":"References"},{"comment":"The user study figure shows only percentages without the number of participants per comparison or any measure of inter-rater agreement; specifying N and a confidence interval would make the figure interpretable.","section":"§4.3, Fig. 6"},{"comment":"The DiffSketcher comparison is qualitative and uses a small set of prompts; this is acceptable as a supplementary illustration, but the main text's claim of 'efficient and effective' text-to-sketch generation should not be read as a quantitative comparison.","section":"Supplementary D"}],"recommendation":"major_revision","confidential_remarks":"The core representation and generator design are plausible and the ablations support the main components, but the evaluation currently does not support the headline 'surpasses state-of-the-art' claim because of the lack of error bars, the metric-validity concern, and the partial user study. I would like to see the strengthened evaluation before acceptance; the method itself appears worth one more revision cycle."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Worth a look if you care about sketch generation or discrete representations for structured output. The genuinely new idea is to treat each stroke as an entity, split its shape from its location, and learn separate vector-quantized codebooks for each. Then a cascaded transformer generates a sketch token by token: first the stroke label, then the shape and location codes. That's a clean design, and the paper shows it works.\n\nThe method is mostly sound. The ablations are informative: removing the VQ representation or the shape/location decoupling degrades FID substantially, which supports the design choices. The code-space clustering (UMAP) showing semantic structure emerging without supervision is a nice result and a good sanity check. The applications—label-, text-, and completion-conditioned generation—are useful and the visual results look plausible.\n\nThe soft spots are in the evaluation, not the method. Table 1 reports point estimates with no error bars or significance tests, and all four metrics are computed in the feature space of an Inception model trained on QuickDraw. That space may not be sensitive to creativity, which is the very thing the paper claims to generate. The authors themselves note that a higher CS can indicate 'simplicity' rather than quality, which undercuts the metric's use in the comparison. A baseline that just returns random training sketches would likely score well on FID/GD, so the table alone does not establish 'surpasses SOTA.' The user study is a step in the right direction but only compares against DoodleFormer (and real data), not DoodlerGAN or SketchKnitter, so it cannot carry the full superiority claim.\n\nThe paper is honest about its limitations: hard cropping, non-exhaustive codebook search, and the teacher-forcing gap are acknowledged in the supplementary, along with detailed network configurations that make the work reproducible in principle. Code is not released, which is a minor strike.\n\nMy take: this is a solid empirical contribution with a representation that could be useful beyond creative sketches. The evaluation needs more rigor before I'd accept the headline claim. I'd send it to review anyway—the idea is worth referee time. If the authors add error bars, a memorization check, and a broader user study, the paper becomes much stronger.","headline":"A clean stroke-level VQ representation with a solid method, but the evaluation needs error bars and a broader user study before 'surpasses SOTA' sticks.","tokens_in":16849,"tokens_out":2888,"would_cite":true,"duration_ms":27880,"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":"VQ-SGen claims that creative sketch generation works best stroke-by-stroke, encoding each stroke as a vector-quantized token and generating sketches with a cascaded autoregressive Transformer that outperforms pixel-based methods on…","keywords":["creative sketch generation","vector quantization","stroke representation","autoregressive Transformer","text-to-sketch","sketch completion","CreativeSketch dataset","discrete codebook"],"falsifier":"A concrete check: re-run the CreativeSketch comparison with an evaluator not trained on QuickDraw (for instance a text-aligned visual encoder or a sketch-specific perceptual metric) and see whether VQ-SGen still beats DoodleFormer on FID; alternatively, run a human creativity-preference study on pairs matched for recognizability to see whether the fine-grained stroke quality the paper claims is perceptible to raters.","tokens_in":15890,"feed_emoji":"✏️","tokens_out":8557,"duration_ms":74976,"temperature":0.7,"pith_summary":"VQ-SGen sets out to show that creative sketch generation should be done at the level of individual strokes rather than pixels or whole parts. The paper proposes a two-stage pipeline: first, it decouples each stroke's shape from its location and encodes both into separate vector-quantized codebooks; second, a cascaded autoregressive Transformer generates a sketch as a sequence of stroke labels, shape codes, and location codes. On the CreativeSketch benchmark, this outperforms existing methods on FID and generation diversity (FID 15.78 vs 17.48 on Creative Birds, 17.61 vs 20.43 on Creative Creatures) and produces sketches with fewer of the blurry regions and broken strokes seen in earlier work. If true, this means a discrete stroke-token representation is a strong basis for sketch generation, and it also opens label-, text-, and completion-conditioned drawing from the same model.","feed_headline":"Stroke tokens beat pixel generation for creative sketches","feed_subtitle":"VQ-SGen cuts FID to 15.78 on Creative Birds and 17.61 on Creative Creatures, with higher diversity.","key_machinery":"The load-bearing object is the vector-quantized stroke representation paired with a cascaded autoregressive Transformer. A stroke is rasterized as a 256x256 image, its axis-aligned bounding box coordinates $(w/2, h/2, x, y)$ define location, and the shape is the stroke translated to image center; two codebooks (each 8192 entries, 512 dimensions in the final configuration) quantize shape and location separately. Two Transformer decoders then predict, in order, the next stroke's label and its shape/location codes, sampling in the discrete code space instead of raw pixels. Training uses reconstruction, codebook, and commitment losses from VQ-VAE, with CoordConv layers and distance-field supervision in the stroke autoencoder. The decoupling forces the shape codebook to specialize on stroke geometry and the position codebook on inter-stroke layout, which the ablations show is what drives the FID improvement (15.78 vs 17.14 without decoupling on Creative Birds).","core_discovery":"The paper's central claim is that a sketch is best modeled as a sequence of stroke entities, each represented by a compact discrete code for its shape and another for its position, and then regenerated autoregressively. VQ-SGen first translates each rasterized stroke so its bounding box is centered, records the box as the location code, learns a VQ codebook for stroke shapes and another for locations, and then trains two cascaded Transformer decoders: one predicts the next stroke's semantic label, the other predicts its shape and position codes conditioned on that label and on all previous strokes. The authors report that this discrete, decoupled representation outperforms pixel-level and point-level baselines on the CreativeSketch dataset, that the learned shape code space clusters by semantic stroke role (beak, wing, tail, eye, body) without semantic supervision, and that the same machinery supports class-label, text, and completion conditioning. The argument is that the VQ representation compresses away redundancy while preserving stroke shape variations, giving the generator a semantically aware, compact space to sample from.","pith_inferences":["A step the authors leave implicit is that the discrete code space could support stroke-level editing or retrieval by swapping or interpolating code indices; the paper demonstrates code interpolation for reconstruction but not as a generation-time control.","Because the label Transformer can be removed with only a small quality drop, the approach is a plausible fit for unlabeled sketch corpora beyond CreativeSketch, including free-hand doodles without part annotations.","A testable extension would be scaling up the codebooks: the paper does not exhaust hyperparameters and reports remaining artifacts like disconnected wings, so larger code spaces might close those gaps.","The reported gains are measured in one fixed feature space; using a different visual encoder could reorder the baselines, which would be a useful sensitivity check for the field."],"forward_implications":["If the claim holds, stroke-token generation replaces pixel-level and part-level generation as the stronger approach on CreativeSketch, with lower FID and higher diversity.","Class-label and text conditioning work by swapping the start token, so one trained generator covers unconditional, conditional, and completion modes without a new architecture.","Sketch completion reduces to seeding the first stroke, and the paper's comparison shows this beats the pixel-level baselines at completing a sketch from an initial mark.","The near-parity of the w/o T_l ablation indicates the method does not depend on stroke labels, so it transfers to unlabeled datasets like the QuickDraw subset used in the application.","The semantic clustering that emerges in the shape code space gives a natural prior for sampling strokes that fit a part role, which the paper suggests can be reused in other sketch-related tasks."],"supporting_citations":[{"why":"Supplies the DoodleFormer baseline that VQ-SGen must beat in the CreativeSketch comparisons.","marker":"[4]"},{"why":"Introduces the CreativeSketch dataset (Creative Birds and Creative Creatures) and the DoodlerGAN part-based baseline, defining the task.","marker":"[8]"},{"why":"Supplies the stroke-point-level SketchKnitter baseline, the third competitor in quantitative and visual comparisons.","marker":"[35]"},{"why":"Provides the vector quantization machinery, codebook, commitment, and reconstruction losses that VQ-SGen adapts to stroke tokens.","marker":"[30]"},{"why":"Provides the autoregressive Transformer decoder architecture used in both generation stages.","marker":"[31]"},{"why":"Source of the single-stroke-as-entity formulation plus the CoordConv and distance-field supervision in the stroke autoencoder.","marker":"[34]"},{"why":"Defines the Inception architecture whose feature space hosts the FID, GD, CS, and SDS evaluation metrics.","marker":"[27]"},{"why":"Supplies the QuickDraw dataset reference used to train the Inception evaluator and for the class-label-conditioned QuickDraw experiment.","marker":"[38]"},{"why":"Source of the idea to decouple a stroke into shape, location, and order embeddings, which VQ-SGen adapts into its decoupled VQ representation.","marker":"[22]"}],"fun_headline_variants":["Sketch strokes as discrete tokens beat pixel generation","VQ stroke codes improve creative sketch generation","Discrete stroke shapes and positions generate better sketches","Stroke-token model tops pixel methods on creative sketches","Quantized stroke representation yields high-fidelity sketches"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the Inception feature space trained on QuickDraw doodles, together with the FID, GD, CS, and SDS metrics computed in it, actually tracks human judgments of creativity and sketch quality; if it does not, the numerical superiority over prior methods would not establish the paper's conclusion.","fun_headline_variants_meta":{"raw":{"variants":["Sketch strokes as discrete tokens beat pixel generation","VQ stroke codes improve creative sketch generation","Discrete stroke shapes and positions generate better sketches","Stroke-token model tops pixel methods on creative sketches","Quantized stroke representation yields high-fidelity sketches"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000168,"raw_usage":{"total_tokens":1264,"prompt_tokens":951,"completion_tokens":313,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":567,"completion_tokens_details":{"reasoning_tokens":242}},"tokens_in":567,"tokens_out":313,"duration_ms":3322,"temperature":1.0,"reasoning_tokens":242,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T13:05:34.991397+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A concrete check: re-run the CreativeSketch comparison with an evaluator not trained on QuickDraw (for instance a text-aligned visual encoder or a sketch-specific perceptual metric) and see whether VQ-SGen still beats DoodleFormer on FID; alternatively, run a human creativity-preference study on pairs matched for recognizability to see whether the fine-grained stroke quality the paper claims is perceptible to raters.","supporting_citations":[{"cited_title":"Doodleformer: Creative sketch drawing with transformers","cited_arxiv_id":null,"evidence_quote":"Supplies the DoodleFormer baseline that VQ-SGen must beat in the CreativeSketch comparisons."},{"cited_title":"Creativ sketch generation","cited_arxiv_id":null,"evidence_quote":"Introduces the CreativeSketch dataset (Creative Birds and Creative Creatures) and the DoodlerGAN part-based baseline, defining the task."},{"cited_title":"Sketchknitter: Vectorized sketch generation with dif- fusion models","cited_arxiv_id":null,"evidence_quote":"Supplies the stroke-point-level SketchKnitter baseline, the third competitor in quantitative and visual comparisons."},{"cited_title":"Neural discrete representation learning","cited_arxiv_id":null,"evidence_quote":"Provides the vector quantization machinery, codebook, commitment, and reconstruction losses that VQ-SGen adapts to stroke tokens."},{"cited_title":"Contextseg: Sketch seman- tic segmentation by querying the context with attention","cited_arxiv_id":null,"evidence_quote":"Source of the single-stroke-as-entity formulation plus the CoordConv and distance-field supervision in the stroke autoencoder."},{"cited_title":"Deep learning for free-hand sketch: A survey","cited_arxiv_id":null,"evidence_quote":"Supplies the QuickDraw dataset reference used to train the Inception evaluator and for the class-label-conditioned QuickDraw experiment."},{"cited_title":"Sketchxai: A first look at explain- ability for human sketches","cited_arxiv_id":null,"evidence_quote":"Source of the idea to decouple a stroke into shape, location, and order embeddings, which VQ-SGen adapts into its decoupled VQ representation."}],"review_version":1}