{"id":"d43f4a17-d13f-4ad5-b610-89b79a7261c9","arxiv_id":"2608.12064","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":8,"one_line_summary":"Draw order is encoded as color in an image, generated by a pretrained diffusion transformer, then decoded into ordered vector strokes whose order follows language instructions.","lead":"A sketch model predicts an order field that says which strokes to draw first, rather than predicting strokes one by one. From an image or text prompt it produces ordered vector strokes, and a text caption can change the draw order without changing the picture.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The abstract's 'for either, it follows text instructions specifying order' is unverified for text-to-sketch: every quantitative order test is image-conditioned derendering, while text-only generation is scored only by CLIP recognition, not order.","rationale":"The reader's CONDITIONAL verdict is appropriate, but I locate the load-bearing weakness differently. The codec concern raised by the reader is substantially mitigated by the order-native decoder: Appendix C shows arc Spearman >= 0.994 and mask IoU >= 0.97 on the decoder ceiling, so the information survives the VAE encode-decode path; the bottleneck is the DiT's ability to render the order field, which is exactly what Table 1 demonstrates at region level. The unaddressed weak point is the text-to-sketch branch: every reported order metric is measured with an input image present. Table 2's CLIP accuracy only proves the image prior's drawing ability survives; it says nothing about whether a purely text prompt that specifies order is followed. The paper's own Section 6 admits within-unit control is absent and part-level tau is 0.46, so the abstract's 'order of drawing' must be scoped to coarse named units. A focused text-to-sketch order-intervention test would determine whether the 'for either' clause holds. Neither this concern nor the reader's changes the verdict from CONDITIONAL: the derendering evidence is strong and honestly reported, but the central claim needs calibration and the text-only order capability needs evaluation.","tokens_in":9960,"tokens_out":11145,"duration_ms":102090,"concrete_test":"Run the order-intervention protocol of Table 1 in text-to-sketch mode: prompt the model with text-only captions that state a part order (e.g., 'a bird: first the head, then the body, then the wings') on an order-annotated test set such as ControlSketch-Part or the in-domain holdout, and compute Kendall tau between predicted stroke order and the stated order, using the model's own HDBSCAN embeddings matched to the named parts to label generated strokes. Compare stated-order vs reversed-order captions with fixed seed; if reversing the caption flips tau sign as in Table 1, the 'for either' claim is supported; if tau stays near zero, the abstract must be limited to derendering and coarse region/part order.","verdict_should_be":"UNCHANGED","load_bearing_attack":"All quantitative evidence for order following comes from the derendering setting, where an input image is present. Table 1 caption interventions, Table 3 granularity funnel, and the external ControlSketch-Part test all condition on an input sketch. For the text-to-sketch branch of the central claim, Section 4.3 reports only CLIP Top-1 recognition on QuickDraw categories; no order instruction is issued and no order metric is computed. The model is said to be trained on 'a text-conditioned generation objective over permuted order' (Section 3.1), so the capability is plausible, but the claim 'for either, it follows text instructions specifying the order of drawing' is not backed by any measured result in the text-only setting. In addition, the abstract's unqualified 'order of drawing' overstates the demonstrated control: Section 5 shows adherence drops to 0.46 at part level and within-unit tau is near zero, so the system follows instructions only at coarse named-unit granularity. Since the central claim is about a general order channel, the missing text-to-sketch order evaluation is the weakest link: if order control does not transfer to generation without an input image, the claim fails for half of the stated capabilities.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper proposes a sketch-order representation and pipeline. It encodes stroke order as an HSV 'codec image' (global arc in hue, within-stroke progress in value), trains a LoRA-adapted latent flow-matching transformer to generate this intermediate from text or image, and finetunes the VAE decoder with new heads that emit a global arc field, foreground mask, and per-stroke instance embeddings. A non-learned vectorizer (HDBSCAN plus nearest-neighbor tracing and RDP simplification) turns these fields into ordered polylines. The model is trained on 47,318 commissioned drawings with programmatic captions that describe a permuted region-level order. The main quantitative results are: (i) a decoder ceiling of Kendall tau around 0.91-0.94 on multi-stroke sets; (ii) caption interventions in derendering that move or reverse stroke order while geometry stays fixed; (iii) region-level instruction adherence tau=0.838, falling to 0.461/0.444 at part/pass level; (iv) an external ControlSketch-Part test with stated-order tau=0.778; and (v) QuickDraw text-to-sketch CLIP recognition of 0.70-0.75 Top-1 without any order evaluation. The paper claims that the system 'can predict an ordered vector sketch from a text description or derender an image into ordered vectors; for either, it follows text instructions specifying the order of drawing.'","tokens_in":10247,"tokens_out":6645,"duration_ms":61605,"significance":"If the order channel works as claimed, this is a novel and useful capability: it makes stroke order a free-form, language-conditioned output variable rather than an implicit generation axis, and it reuses a strong pretrained image prior. The evaluation is more careful than typical for this area: the authors separate decoder ceiling from end-to-end loss, include caption-deletion, recorded-order, and reversed-order controls with geometry held fixed, report a condition funnel, and validate on an externally annotated dataset whose part annotations come from a different pipeline. These controls give me moderate confidence that language genuinely influences coarse stroke order in the derendering setting. The significance is reduced, however, by two gaps: the text-to-sketch branch of the central claim has no order evaluation, and the demonstrated order control is coarse (named units only), with the paper itself stating that 'below named units there is no control at all' (Section 6). The core idea and evaluation design are solid enough to warrant revision rather than rejection.","major_comments":[{"comment":"The abstract's claim that 'for either, it follows text instructions specifying the order of drawing' is not supported for the text-to-sketch branch, because every quantitative order test in the paper—Table 1, Table 3, and the ControlSketch-Part evaluation—conditions on an input sketch, while Section 4.3 evaluates text-only generation only with CLIP recognition over 50 QuickDraw categories and never issues or measures an order instruction. Please add a text-to-sketch order experiment (for example, generating with explicit order captions and scoring part-order adherence against an annotation or human judgment), or narrow the claim to image-conditioned derendering.","section":"Section 4.3, Table 2"},{"comment":"The phrase 'follows text instructions specifying the order of drawing' overstates the measured precision. Region-level instruction adherence is high (tau=0.838), but part-level adherence drops to 0.461, pass-level to 0.444, and within-unit residual tau is near zero (0.177/0.134 on holdout; 0.058/0.068 on ControlSketch-Part). Section 6 itself concedes that 'below named units there is no control at all.' The claim should be qualified as following coarse named-unit order, not arbitrary or stroke-level order.","section":"Abstract, Section 5, Table 3"},{"comment":"All headline order numbers are n=100 point estimates with no confidence intervals and no description of how the 100 drawings per dataset are selected (random, first N, or stratified), and the Wilcoxon tests in Appendix B are reported only for one comparison. Because the paper's central quantitative claims—especially the part-level and ControlSketch-Part numbers—are used to calibrate how much instruction following the system achieves, the authors should report bootstrap confidence intervals or per-sample distributions and state the selection procedure.","section":"Section 4, Tables 1, 3, 4"}],"minor_comments":[{"comment":"The holdout path-count ratio of 12.53 conflicts with Section 6's statement that recovered paths run 1.7-2.5 times the true stroke count; please reconcile or correct the table entry.","section":"Table 4"},{"comment":"The sentence 'The pixel branch is architecturally unchanged, while training jointly with the new head' is ambiguous about whether the reconstruction branch is also finetuned; clarify which weights are updated.","section":"Section 3.3"},{"comment":"HDBSCAN clustering over the eight-dimensional embedding channels is a key step for segmentation, but the appendix gives no hyperparameter values or sensitivity analysis for HDBSCAN or the capping radius in Algorithm 1; please add these to the appendix.","section":"Section 3.4 and Algorithm 1"},{"comment":"The primary order metric is called Kendall tau, but the paper does not state whether this is tau-a or tau-b; ties can arise from the vectorizer's duplicated arc values, and the choice affects comparability across tables.","section":"Section 4"},{"comment":"The 'pass' level of instruction is referenced in Table 3 but not defined until Section 6; define it earlier or add a pointer.","section":"Section 5"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is an industrial preprint with no public code or data release, so reproducibility is limited despite the careful evaluation design. I would ask the editor to treat the missing text-to-sketch order evaluation as a mandatory revision rather than a suggestion, because the abstract's central claim names both branches. The paper is not a reject: the derendering evidence is strong, the external ControlSketch-Part validation is a real strength, and the coarse-grained order-control results are useful even if the abstract overreaches."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You should know this paper is a real systems contribution, not a repackaging. The order-as-color codec and the order-native decoder are new, and the paper does the legwork to show the idea works: it separates decoder ceiling from end-to-end, includes random/reversal controls, and reports an external part-annotated test. The decoder ceiling numbers are strong (arc Spearman ≥0.99, matched Kendall τ 0.91–0.94), and the caption interventions in Table 1 make a convincing case that the caption, not the input image, is the order channel in derendering. The limitations section is unusually honest and should be kept in any revision.\n\nThe soft spots are real but not fatal. The stress-test is right: every quantitative order result comes from image-conditioned derendering. Table 1, Table 3, and the ControlSketch-Part test all condition on an input sketch. The text-to-sketch branch of the central claim is supported only by CLIP recognition on QuickDraw categories with no order instruction issued, so the abstract's 'for either, it follows text instructions specifying the order of drawing' is not backed by a measured order metric in the generation setting. That is a gap, not a refutation—the model is trained with text-conditioned permuted-order objectives, so the capability is plausible—but the claim should be narrowed or the experiment run.\n\nI also want to flag, proportionately: within-unit order agreement is near zero under every instruction, so 'order control' is actually coarse region-level control, not stroke-level control. That is a genuine scope limitation that the abstract obscures. The paper reports no error bars on its central n=100 point estimates, and the dataset is private, so independent verification is limited. Fragmentation (path counts 1.7–2.5× true strokes) is acknowledged and quantified.\n\nNone of this undermines the core idea: the paper convincingly shows that a pretrained image model can be adapted to produce ordered vector sketches and that order can be steered by text in the derendering setting. The missing piece is the equivalent demonstration for text-to-sketch, and the paper would be much stronger with a focused experiment on that.\n\nThis deserves serious peer review. I would send it out, with a request for the text-to-sketch order test, error bars, and a more measured abstract. I'd bring it to reading group, and I'd cite the codec idea if I were working on sketch generation.","headline":"A genuinely new order-as-color system with careful derendering evaluation, but the headline claim overreaches: text-to-sketch order is never quantitatively tested.","tokens_in":10785,"tokens_out":2257,"would_cite":true,"duration_ms":21300,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Drawing order can be encoded as a color field and read back as text-ordered strokes.","keywords":["sketch generation","stroke order","order-as-color codec","image derendering","text-conditioned drawing","latent flow matching","diffusion transformer","vectorization"],"falsifier":"Take a set of drawings whose true stroke-level order is known and issue text instructions that name individual strokes or atomic units in a specific order. If the resulting Kendall tau stays near zero, while region-level instructions reach 0.8, the claim of following text instructions for drawing order is true only at region granularity. A cleaner test: generate the same sketch twice under 'draw top to bottom' and 'draw bottom to top', then check whether every matching stroke's relative order flips; any stroke whose order does not invert marks the control boundary.","tokens_in":9712,"feed_emoji":"🎨","tokens_out":7106,"duration_ms":64875,"temperature":0.7,"pith_summary":"This paper claims that the order in which strokes are drawn can be treated as a 2D image field instead of a sequence: it encodes global drawing progress into hue and within-stroke progress into brightness, generates or derenders that encoded image with a pretrained latent flow-matching transformer, and reads the field back out with a retrained decoder to assemble polylines sorted by draw order. The system thereby gives a text-conditioned channel for stroke order: captions state which regions or parts come first, and changing only the caption changes the order while leaving the geometry fixed. The clearest evidence is that reversing a caption flips Kendall tau from 0.373 to -0.267 with unchanged mask IoU, and that on an external part-annotated benchmark the model reaches 0.778 stroke-order Kendall when the stated part order is given. In the paper's own tests, order control degrades with instruction granularity: region-level instructions are followed reliably, part-level adherence is 0.46, and stroke order inside a named unit is near zero. The contribution is a representation-level inversion: order is no longer a generation axis but a property of a generated image that can be read back.","feed_headline":"Sketching order rides a color field through a diffusion model","feed_subtitle":"Draw order is painted into hue and brightness, survives generation, and comes back as replayable strokes.","key_machinery":"The load-bearing object is the order-as-color codec of Eq. (1): each inked pixel's hue is $H = a\\cdot 342/360$, where $a\\in[0,1]$ is the accumulated arc length from the first stroke to that pixel divided by total arc length; saturation is 1 on ink and 0 on background; and value is $V = 1-u/2$, where $u$ is the arc fraction within the current stroke. This single 8-bit RGB image lets a pretrained latent flow-matching transformer generate or derender the order field inside its native image latent space. A retrained decoder with a pyramid CNN head emits the predicted global arc field, a foreground mask, and eight-dimensional stroke-instance embeddings; HDBSCAN clusters the embeddings, a nearest-neighbour walk recovers polylines, and the global arc field sorts clusters into the final stroke order. The codec is what converts a sequential ordering problem into a spatial image-generation problem, so text can steer order through the diffusion model and the decoder can read it back.","core_discovery":"The central discovery is that vector stroke order can be carried lossily through an image-generative model. The paper shows that if every inked pixel is colored by a global arc fraction and a within-stroke arc fraction, then a frozen image encoder, a diffusion transformer, and a finetuned decoder can recover enough of that field to reconstruct ordered vectors with Kendall tau 0.91–0.94 at the decoder ceiling and 0.449 with recorded-order captions end to end. Order is an independent controllable dimension: geometry metrics stay flat while order swings from positive to negative when the text instruction is reversed, and the model continues to generate recognizable sketches because it retains the base model's world knowledge. The authors state the proposal as: instead of drawing strokes in order, predict a 2D field that defines the order, then vectorize and sort by it.","pith_inferences":["Extension beyond the paper: because the order field is a continuous scalar independent of stroke geometry, the same codec could convert offline handwriting or assembly video into ordered trajectories, provided a dataset with recorded ground-truth order exists.","The caption-reversal result suggests a testable design principle: when a language instruction conflicts with an image condition, the diffusion model treats the text as the stronger order prior; measuring this conflict systematically could guide conditioning for other generative tasks.","A direct follow-up experiment would train the same architecture with permutations at the stroke or point level rather than the region level; if the codec survives, the paper's own limitation list predicts within-unit tau should rise from near zero.","If per-point order labels were available, the global arc field could be replaced by any learned monotone embedding, turning the method into a general order-as-image representation for sequential data."],"forward_implications":["Because text, not geometry, carries the order, the same drawing can be replayed in multiple orders from one model by changing only the caption.","Order control is bounded by the granularity of the training permutation: coarse units are reliable, part-level adherence is about 0.46, and within-unit stroke order is essentially uninstructable.","The decoder ceiling (Kendall tau 0.91–0.94 on four datasets) shows that the vectorization path is not the bottleneck; losses in end-to-end order come from generation given the caption.","The model can derender an input image into replayable ordered vectors while retaining the base model's open-vocabulary text-to-sketch ability, as measured by CLIP recognition of generated QuickDraw categories."],"supporting_citations":[{"why":"Supplies the pretrained latent flow-matching transformer and VAE whose latent space carries the codec image.","marker":"[Wu et al., 2025]"},{"why":"Defines the latent flow-matching objective used to train the generator on the encoded order field.","marker":"[Lipman et al., 2023]"},{"why":"Provides the low-rank adaptation method used to finetune the diffusion backbone.","marker":"[Hu et al., 2022]"},{"why":"Provides the external part-annotated ControlSketch-Part benchmark used to measure instructed part order.","marker":"[Du et al., 2026]"},{"why":"Supplies the Creative Birds and Creative Creatures out-of-domain test sets.","marker":"[Ge et al., 2021]"},{"why":"Supplies the FS-COCO out-of-domain test set.","marker":"[Chowdhury et al., 2022]"},{"why":"Supplies the QuickDraw categories used for text-to-vector recognition evaluation.","marker":"[Jongejan et al., 2016]"}],"fun_headline_variants":["Order field predicts sketch stroke sequence","Draw order rides a color field through diffusion","Text controls sketch order via a 2D field","Reversible order field for vector sketches","Diffusion model sorts strokes by painted field"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The assumption that carries the whole system is that an 8-bit color map of draw order can pass through the lossy image encoder and diffusion model with enough of its fine-grained hue and brightness structure intact for the decoder to sort strokes correctly; if the image prior smooths away that structure, the order signal is gone.","fun_headline_variants_meta":{"raw":{"variants":["Order field predicts sketch stroke sequence","Draw order rides a color field through diffusion","Text controls sketch order via a 2D field","Reversible order field for vector sketches","Diffusion model sorts strokes by painted field"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000158,"raw_usage":{"total_tokens":1154,"prompt_tokens":805,"completion_tokens":349,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":421,"completion_tokens_details":{"reasoning_tokens":284}},"tokens_in":421,"tokens_out":349,"duration_ms":3913,"temperature":1.0,"reasoning_tokens":284,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T00:17:48.989458+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a set of drawings whose true stroke-level order is known and issue text instructions that name individual strokes or atomic units in a specific order. If the resulting Kendall tau stays near zero, while region-level instructions reach 0.8, the claim of following text instructions for drawing order is true only at region granularity. A cleaner test: generate the same sketch twice under 'draw top to bottom' and 'draw bottom to top', then check whether every matching stroke's relative order flips; any stroke whose order does not invert marks the control boundary.","supporting_citations":[],"review_version":1}