{"id":"ba837d6b-e7b4-4c9f-ac8b-5ac53c48dcca","arxiv_id":"2501.03992","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"NeuralSVG encodes a text-prompted vector graphic as the weights of a small MLP, using nested dropout so the first learned shapes are semantically meaningful and editable.","lead":"NeuralSVG generates vector graphics from text by training a small neural network to output each shape, with a dropout-based ordering trick that makes early shapes the meaningful skeleton. The result is a vector graphic whose layers can be edited, recolored, or reshaped at inference time.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"No random-dropout control isolates whether nested dropout, rather than any dropout, causes the reported ordering; the cumulative CLIP metric also mirrors the training objective.","rationale":"The reader's weakest_assumption correctly identifies that the ordering claim is probed only indirectly by cumulative CLIP curves and lacks direct editability metrics. My concern sharpens this: the cumulative CLIP metric is not simply indirect, it is partially circular, because nested dropout explicitly trains prefixes to match the prompt under a CLIP-aligned diffusion prior. The missing random-dropout control means the observed effect could be generic dropout regularization rather than the nested, prefix-ordered mechanism that the paper claims. This does not invalidate the paper's results, but it means the central mechanism is underdetermined by the provided evidence. The paper does give credit where due: the architecture is simple and clearly described, the LoRA fine-tuning details are in the appendix, the qualitative comparisons to open-source baselines are extensive, and the sketch-generation extension is a nice contribution. A concrete random-dropout ablation is a feasible and decisive experiment that would settle whether the 'ordered' part of the claim is real. Since this concern is addressable and does not contradict the qualitative findings, the appropriate verdict remains conditional acceptance, matching the reader's verdict.","tokens_in":17644,"tokens_out":4053,"duration_ms":45568,"concrete_test":"Re-train NeuralSVG with a random-dropout control: at each iteration, with probability 0.7 render all 16 shapes; otherwise sample a truncation count t from the same exponential distribution (temperature 3) and render a uniformly random subset of exactly t shapes (not the first t). Keep all other hyperparameters identical. Then compute (a) cumulative CLIP similarity curves when rendering the learned shapes in index order (as in Figures 9 and 11) and (b) a direct editability metric, e.g., the change in full-SVG CLIP or pixel IoU when deleting each individual shape, or a human rating of whether single shapes are identifiable. If the random-dropout model achieves prefix-CLIP curves comparable to the nested-dropout model, the paper's ordered-decomposition claim is not supported, because any dropout suffices.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim is that the nested-dropout variant (Section 4.3) encourages each learned shape to have a meaningful and ordered role in the composition, enabling truncation-based editing and layered output. The evidence offered is (i) qualitative progressions in Figures 5, 15, 16, and (ii) cumulative CLIP curves in Figures 9 and 11, plus the no-dropout ablation. This evidence has a confound: the cumulative CLIP curve measures exactly what the training procedure optimizes. With nested dropout, each rendered prefix is a random subset seen during SDS training, so the diffusion prior (and the CLIP-like signal inside SDS) is applied directly to prefixes. High prefix CLIP similarity therefore reflects successful optimization of the training objective, not necessarily semantic decomposition or editability. More importantly, the ablation only compares nested dropout against no dropout. It does not compare against a random-dropout baseline where the same number of shapes is dropped uniformly at random instead of as a prefix. If any dropout—regardless of ordering—improves prefix recognizability, then the reported cumulative CLIP improvement cannot be attributed to the 'ordered' component of the mechanism. The specific claim that Rippel et al.'s nested-dropout ordering guarantee transfers to SDS-trained shape sets remains untested, because the design does not isolate the ordering from the dropout itself. This is the load-bearing weakness in the argument.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"NeuralSVG proposes to represent an SVG as the output of a small MLP that maps a shape index (optionally conditioned on background color or aspect ratio) to a closed Bezier shape with fill color. The network is trained by Score Distillation Sampling (SDS) from a Stable Diffusion model, using a 'nested dropout' scheme that at each iteration truncates the shape set at a random index and renders only the prefix. The paper claims that this dropout encourages an ordered, semantically meaningful layering of shapes, so that early shapes capture the main structure and later shapes add detail, and that the implicit representation enables inference-time controls such as background color and aspect ratio. The method is compared with VectorFusion, SVGDreamer, NIVeL, and Text-to-Vector using CLIP-based metrics and qualitative visuals, plus ablations and a sketch-generation extension.","tokens_in":17948,"tokens_out":5613,"duration_ms":51166,"significance":"The goal of producing compact, editable, layered text-to-vector output is well motivated, and the use of nested dropout to impose ordering is a creative and simple idea that could be useful beyond this setting. The paper is clearly written, and the supplementary visual results show that the method produces visually appealing SVGs and plausible sketches with an ordered progression when rendered with progressively more shapes. If the ordering claim could be rigorously established, this would be a meaningful step toward practical layered vector generation. However, the central evidence for the ordering claim is currently indirect and partially confounded with the training objective, and the quantitative comparison against high-shape-count baselines is less favorable than the text suggests. The strengths of the paper are the clean formulation, the single-stage training, and the demonstrated inference-time controls, though the latter are only shown qualitatively.","major_comments":[{"comment":"The claim that the proposed dropout encourages an 'ordered' representation is not isolated from the effect of dropout in general. The ablation compares nested dropout only to no dropout. Since any dropout that randomly omits shapes during SDS training might improve the recognizability of prefixes (for example, by regularizing early shapes or by preventing overfitting to the full composition), the reported cumulative CLIP gains cannot be attributed to the ordering (prefix) structure. Add a control that drops the same number of shapes uniformly at random rather than as a prefix; if the two curves are similar, the ordering claim is unsupported. In addition, the Rippel et al. nesting guarantee for hidden-unit representations does not automatically transfer to SDS-trained shape sets, so a direct check of semantic identifiability of early shapes (e.g., human evaluation or VLM-based part labeling) is needed.","section":"Section 4.3, Eq. (7), Figures 10-11"},{"comment":"The statement that 'all methods achieve comparable CLIP scores' when VectorFusion uses 64 shapes and SVGDreamer uses 256 shapes is inaccurate: NeuralSVG's R-Precision is 67.18, well below 83.46 and 85.03 for these baselines, a gap of roughly 16-18 points. The text should either report these numbers accurately and emphasize that the advantage is in structure/editability rather than raw CLIP alignment, or provide a quantitative editability/structure metric (e.g., user study, number of meaningful semantic segments) to support the 'outperforms' claim.","section":"Section 5.2, Table 1"},{"comment":"The cumulative CLIP curve is entangled with the training objective because with nested dropout every rendered prefix is exactly one of the training-time renderings. The SDS loss (Eq. 2) applied to these prefixes means the diffusion model's CLIP-derived signal directly optimizes prefix recognizability. Thus high cumulative CLIP similarity partly reflects successful optimization of the training loss, not necessarily an emergent, editable semantic order. Report at least one independent measure that does not use CLIP or the SDS signal, such as a human study of the recognizability/meaning of each prefix, or a per-shape semantic labeling with a VLM not involved in training.","section":"Section 5.2, Figure 11"}],"minor_comments":[{"comment":"There is a typo 'ˆ𝑐𝑖𝑖' (double subscript i) in the definition of the color output.","section":"Section 4.2, Eq. (5)"},{"comment":"The notation 'S𝑡𝑟' should be 'S_{t_r}' to be consistent with the text; also clarify whether the truncation index is inclusive or exclusive (i.e., whether shape i < t_r or i ≤ t_r).","section":"Section 4.3, Eq. (7)"},{"comment":"The sampling of the truncation index 'sampled from an exponential distribution with a temperature value of 3' is ambiguous; specify how the discrete distribution over {1,...,16} is constructed and how the 0.7 probability of using all 16 shapes interacts with it.","section":"Appendix, Training Scheme"},{"comment":"The name 'NiVEL' is sometimes spelled 'NiVEL' and sometimes 'NIVeL'; standardize the spelling.","section":"Section 5.5 and throughout"},{"comment":"The word 'Additioanl' is a typo for 'Additional'.","section":"Figure 24 caption"},{"comment":"The description says the adapter was pretrained on 'a dataset of vector art images' generated using 'Simple Vector Flux LoRA'; clarify the provenance and whether this is a dataset of generated images or of real vector artwork, and provide details of the prompt variations.","section":"Appendix, LoRA Fine-Tuning"},{"comment":"The paper does not state whether the code and trained models will be released; please include an availability statement to support reproducibility.","section":"Reproducibility"}],"recommendation":"major_revision","confidential_remarks":"The main gap is the missing random-dropout control and the reliance on CLIP-based metrics that mirror the SDS objective. If the authors can supply a random-dropout ablation and an independent semantic-ordering evaluation, the paper would be much stronger. Given the otherwise clear writing and interesting idea, I recommend major revision rather than rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: NeuralSVG is a legitimate, well-executed integration paper. It does not break new theoretical ground, but the combination of an index-conditioned MLP that emits Bezier path parameters directly, nested-dropout regularization for ordering, and inference-time conditioning (background color, aspect ratio) is new and practically useful. The paper is clearly written, the method is specified in enough detail to re-implement, and the supplementary material is unusually generous with qualitative results.\n\nWhere it holds up: the ordering claim is supported by the truncation experiments in Figures 5, 15, and 16 — keeping the first 1, 4, 8, 12 shapes gives a sensible progressive build-up of the scene. That evidence does not depend on CLIP and is fairly convincing. The color palette and aspect ratio controls also look real, not just cherry-picked. The authors are honest that NIVeL and Text-to-Vector are closed-source and comparisons with them are visual only.\n\nWhere it is soft: the dropout ablation compares nested dropout against no dropout, but not against a random-dropout control where the same number of shapes is dropped uniformly at random. Without that control, you cannot fully attribute the cumulative CLIP improvement to the ordered component of the mechanism. The cumulative CLIP metric itself also overlaps with the SDS training signal, since prefixes are what the model is optimized on. That said, the qualitative truncation results mitigate this concern — random dropout is unlikely to produce such clean progressions. Other weaknesses are standard for the area: no code release, no error bars, and CLIP metrics as the sole quantitative measure. None of these are load-bearing.\n\nBottom line: the paper deserves a serious referee. I would send it to peer review with two firm requests: add a random-dropout ablation and release code. It is a useful contribution for anyone working on text-to-vector generation or editable vector representations, and it advances the practical state of the art without overclaiming.","headline":"NeuralSVG is a solid, well-executed integration paper with a genuinely useful ordering mechanism; the missing random-dropout control keeps the central claim from being fully proven, but it deserves peer review.","tokens_in":18465,"tokens_out":2156,"would_cite":true,"duration_ms":21908,"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":"NeuralSVG packs an entire SVG into one small MLP, learned from a text prompt.","keywords":["text-to-vector","implicit neural representation","vector graphics","Score Distillation Sampling","nested dropout","layered SVG","diffusion models","inference-time control"],"falsifier":"Render only the first four shapes from a trained 16-shape NeuralSVG and ask annotators to name the subject; if early-only renders are no more identifiable than a random four-shape subset, the ordering claim fails. A second check is to take one early shape, change its color, and see whether the edit stays semantically local; if early shapes are not individually editable, the layered-structure promise is broken.","tokens_in":17480,"feed_emoji":"🎨","tokens_out":5232,"duration_ms":45022,"temperature":0.7,"pith_summary":"The paper argues that vector graphics can be generated from text by treating the whole SVG as the weights of a small neural network, rather than as a list of independently optimized paths. It claims that this implicit representation produces fewer, cleaner, and more meaningful shapes than existing text-to-vector systems, and that a dropout-based ordering trick gives the shapes a built-in layer structure: early shapes carry the coarse scene, later shapes add detail. If true, the same trained network can be queried for a compact, editable SVG, and can be reused to recolor the scene or change its aspect ratio at inference time without retraining. The paper also shows that the same framework generates ordered sketches by switching to open strokes.","feed_headline":"One small MLP holds an entire editable SVG, trained from a text prompt","feed_subtitle":"A single learned representation supports partial renders, recoloring, and aspect-ratio changes without retraining.","key_machinery":"The central object is the implicit shape-index MLP: shape indices $\\{1,\\ldots,n\\}$ pass through random Fourier features and two parallel branches that output 12 control points and an RGB color per shape, aggregated by a differentiable rasterizer. The load-bearing mechanism is nested dropout on the shape index: sampling a truncation $t_r$ and rendering only $\\{p_i,c_i\\}_{i