{"id":"aa05affb-d5f8-4f11-9beb-72388972f262","arxiv_id":"2505.10558","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":7,"one_line_summary":"A two-stage text-to-vector diffusion pipeline distills styles from customized image diffusion models, enabling feed-forward generation of SVGs in user-defined styles.","lead":"This paper presents a two-stage pipeline for generating vector graphics (SVGs) in custom styles from text prompts. It trains a diffusion model on black-and-white icons, then distills styles from customized image-generation models to enable fast, style-consistent vector output.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Frozen path VAE trained on black-and-white icons may not decode colored/decorative styles; no reconstruction evaluation supports this assumption.","rationale":"The reader's weakest_assumption is precisely the same load-bearing concern I identify: the pretrained path VAE from T2V-NPR, trained on black-and-white FIGR-8-SVG icons, must encode and decode colored, decorative, and structurally complex paths after style customization. This is not a peripheral quality issue; it is a necessary condition for the central claim that the pipeline can generate SVGs in custom styles in a feed-forward manner. If the VAE cannot represent stylistic details, the image-level loss in Stage 2 cannot push the model toward valid stylized outputs, and the pipeline collapses. The paper provides no direct evidence that the VAE reconstructs out-of-distribution stylized SVGs well; all reported metrics either use the same VAE (Path FID) or are image-based similarity scores that do not isolate VAE decoding fidelity. The self-training loop in Section 5.2 adds further risk because the model is trained on its own generated s0^g, potentially reinforcing decoder weaknesses. Given the absence of code, data, and a VAE reconstruction check, the reader's CONDITIONAL verdict is appropriate and should not be changed. I would accept the paper only after the VAE capacity concern is addressed with a concrete reconstruction evaluation on stylized SVGs and, ideally, a demonstration that generated stylized SVGs can be re-encoded and decoded without loss of style.","tokens_in":14893,"tokens_out":3055,"duration_ms":31071,"concrete_test":"Collect 50-100 colored, stylized SVGs from the same sources used for style exemplars (SVGRepo, iconfont, Freepik). Encode each with the pretrained path encoder, decode back to paths, re-render, and compute LPIPS and Chamfer distance between the original rendered image and the reconstruction. If mean LPIPS is comparable to reconstruction quality on FIGR-8-SVG, the VAE is not the bottleneck; if reconstruction error is significantly higher on stylized inputs, the central claim fails for styles outside the VAE's training domain. Report qualitative examples alongside the numbers.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The entire Stage 2 style distillation rests on the pretrained path VAE from T2V-NPR (Section 4.1). The decoder is frozen; only the diffusion model is fine-tuned. The image loss in Eq. 3 backpropagates through the fixed decoder and differentiable rasterizer. If the VAE latent space cannot represent the colored, gradient, or ornamented paths typical of customized styles, then no matter how well the diffusion model fits the customized image distribution, the final decoded SVG will be structurally or stylistically wrong. No experiment in the paper measures VAE reconstruction on stylized or colored SVGs. The Path FID metric in Section 6 instead scores generated latents against the same black-and-white FIGR-8-SVG distribution on which the VAE was trained; this metric cannot reveal a representational bottleneck for out-of-distribution styles. Moreover, because Stage 2 training data are self-generated (s0^g comes from the current T2V model), the model may settle on latents that satisfy the image loss pointwise but lie off the VAE manifold, producing decoding artifacts at inference. This is a load-bearing assumption for the central feed-forward style customization claim.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a two-stage pipeline for text-to-vector (T2V) generation with customizable styles. In Stage 1, a path-level diffusion transformer is trained on the black-and-white FIGR-8-SVG dataset. Each SVG is represented as a sequence of path-latent vectors obtained from a pre-trained path VAE from the authors' prior T2V-NPR work, together with per-path color and transform parameters. In Stage 2, the T2V model is customized by distilling from fine-tuned text-to-image diffusion models: for each style, a DreamBooth-LoRA and a ControlNet-equipped Stable Diffusion model generate customized images conditioned on Canny edges of the current model's renderings, and the T2V model is fine-tuned with an image-space MSE loss plus the original diffusion loss. The authors report quantitative comparisons on Path FID, style alignment, aesthetic score, and text alignment, a user study, and qualitative results, and claim this is the first feed-forward T2V model that supports style customization from a few exemplar SVGs.","tokens_in":15144,"tokens_out":4089,"duration_ms":41819,"significance":"If the central claim holds, the work is a useful step: it combines the structural regularity of feed-forward vector diffusion with the customization power of text-to-image diffusion models, and the idea of distilling style through rendered image pairs is sensible. The paper provides extensive qualitative comparisons, a user study, and ablation studies, which are strengths. The two-stage design is clear and the claim of feed-forward customized generation is concrete and falsifiable. However, the evaluation as reported does not yet establish the claim to the standard required by a major venue, because the main vector-level metric is computed in the latent space of the same path VAE that the model itself uses, and because the frozen decoder's ability to represent stylized, colored, and decorated paths is asserted rather than demonstrated.","major_comments":[{"comment":"The Path FID metric is computed by encoding generated SVGs and ground-truth FIGR-8-SVG paths with the path VAE from T2V-NPR, which is exactly the latent representation used by the proposed T2V model. This makes the metric partially self-referential: a model that produces latents close to the training distribution will score well even if its decoded SVGs are visually poor, while optimization-based baselines that are not trained in this latent space are systematically disadvantaged. In addition, Table 1 (and Table 2) report no error bars, confidence intervals, or multiple-seed results, so it is unclear whether the differences among Potrace, GPT-4o, T2V-NPR, and Ours are statistically meaningful. I recommend reporting an image-space FID on rendered SVGs, an independent vector-quality metric, and variance estimates over test prompts and generation seeds.","section":"Section 6, Table 1"},{"comment":"The whole pipeline rests on a pre-trained path VAE/decoder from T2V-NPR that was trained on black-and-white FIGR-8-SVG icons, but the style-customization objective in Eq. (3) trains the diffusion model to produce latents whose renderings match colored, decorated customized images. The paper provides no reconstruction evaluation showing that this VAE can encode and decode stylized paths, including colored fills, ornaments, and complex silhouettes. If the decoder's latent space cannot represent such structures, the image loss is unsatisfiable and the feed-forward outputs will be stylistically wrong regardless of how well the diffusion model fits the image distribution; the authors' own failure case in Figure 8(b) is consistent with this risk. Please add a quantitative encode-decode reconstruction study on the style exemplars and on held-out stylized SVGs, reporting geometric and color fidelity, and discuss how training on self-generated s0^g avoids drifting off the VAE manifold.","section":"Sections 4.1 and 5.2"},{"comment":"The Style Alignment metric, measured as CLIP image cosine similarity to the style reference, is not a reliable indicator of customization quality in this comparison: the vector-fine-tuned VecFusion baseline scores highest (0.726) precisely because it overfits and reproduces the exemplars, while the qualitative results show that it fails to adapt to the text prompts. Reporting style alignment without its trade-off against text alignment is misleading; at minimum the authors should present per-method paired results or a combined metric that penalizes overfitting, and should report the same variance information requested for the other metrics.","section":"Section 6, Table 1"}],"minor_comments":[{"comment":"The heading reads 'Cusomization of T2I Generation' and should be corrected to 'Customization of T2I Generation'.","section":"Section 2.3 heading"},{"comment":"The directional arrows attached to the metric names are not explained; please state in the caption which direction is better for each metric.","section":"Table 1"},{"comment":"Equation (3) writes an L2 norm without squaring while the text calls the loss an MSE; please make the notation consistent by writing the squared L2 norm or clarifying the intended loss.","section":"Equation (3)"},{"comment":"The text says the diffusion loss is defined on the predicted denoised tensor \\hat{s}^g_0, but Eq. (1) defines the diffusion loss on the original SVG tensor s0; please clarify the exact training target for the Stage 2 diffusion loss and whether gradients flow through the reparameterization in Eq. (2).","section":"Section 5.2"},{"comment":"The user study reports preferences from 30 participants on 20 prompts but gives no statistical significance test; please add a significance test or bootstrap confidence intervals for the preference percentages.","section":"Section 6.2 and Figure 6"},{"comment":"The conclusion's admitted limitations, especially the 'inadequate style capture' for complex references in Figure 8(b), should be reflected in the abstract's unqualified claim of 'high-quality and diverse' custom-style generation; please qualify the claim or provide evidence on the failure rate.","section":"Section 7"}],"recommendation":"major_revision","confidential_remarks":"The evaluation depends heavily on the authors' own prior T2V-NPR path VAE without an independent check of its reconstruction fidelity for stylized, colored SVGs. This is a legitimate technical concern rather than a reason to reject, but the requested reconstruction study and the de-biasing of the Path FID metric are essential before the central claim can be accepted. I also note that the paper's headline claim of being the first feed-forward style-customizable T2V model is not the issue; the issue is whether the evidence supports it."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is the first feed-forward text-to-vector model I've seen that handles style customization, and the two-stage trick — train a path-level T2V diffusion model on plain icons, then distill style from a DreamBooth-LoRA-tuned SD model through image loss with ControlNet keeping the structure — is a sensible way to get there. The qualitative results are genuinely impressive: the styled icons look coherent and structurally regular, and the user study, while small, backs the main claim. The authors also state limitations honestly.\n\nWhere it gets shaky is the vector-level evaluation. Path FID compares generated latents to ground-truth latents in the same path-VAE space that the model itself uses (from their T2V-NPR paper). That cannot certify vector quality; it partly measures how well the model stays in its own prior. The table has no error bars or significance tests, so the margin over GPT-4o and T2V-NPR is hard to trust. A cheap fix would be reconstruction and human-editing metrics, plus reporting variance.\n\nThe stress-test worry about the frozen path VAE is real but slightly overstated. Color is carried outside the VAE latent (C_i is separate), so the claim that colored/decorative styles cannot be represented is too strong. The geometry of ornate shapes is still a legitimate risk, and there is no experiment measuring VAE reconstruction on stylized SVG shapes. Since the whole second stage leans on that decoder, I'd want one reconstruction plot before believing the pipeline generalizes to arbitrary styles. Relatedly, Stage 2 uses the model's own generations as training seeds; the diffusion loss may keep it on-manifold, but this is another reason external evaluation matters.\n\nCitation pattern is fine, though the paper leans on T2V-NPR for the VAE and representation — that is a dependency, not a flaw. Missing SVGFusion as a baseline is a minor gap. No code or data release, which weakens reproducibility.\n\nOverall: I think the core claim probably holds — feed-forward style customization for SVGs is a real capability and this is a practical route to it. The evaluation is the weak link, not the method. I would send this to referees, but ask them to nail down the Path FID issue and demand a reconstruction check for the VAE. If the authors add that, it's a solid paper.","headline":"A plausible and useful two-stage recipe for style-customized text-to-vector generation, with a self-referential Path FID and an untested frozen path VAE as the two things to fix before trusting the numbers.","tokens_in":15669,"tokens_out":3178,"would_cite":true,"duration_ms":30186,"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":"Text-to-vector generation can learn a custom SVG style from a few example files and apply it to arbitrary text prompts in a single feed-forward pass.","keywords":["Scalable Vector Graphics","SVG generation","text-to-vector generation","style customization","diffusion model","path-level representation","image diffusion priors","feed-forward generation"],"falsifier":"Take colored, gradient-filled, or multi-layer SVG style exemplars, pass them through the path-level encoder and decoder used by the pipeline, and measure how much of the style survives reconstruction; if decoded SVGs lose color, decoration, or layer structure enough that the style becomes unrecognizable, the pretrained latent space is the bottleneck and the feed-forward customization claim fails for those styles.","tokens_in":14571,"feed_emoji":"🎨","tokens_out":7175,"duration_ms":62696,"temperature":0.7,"pith_summary":"The paper claims that a text-to-vector (T2V) model can learn a custom visual style from just a few example SVG files, and that after learning it can generate new SVGs in that style from any text prompt in a single forward pass. The proposed pipeline separates the problem into two stages: first train a diffusion model that generates well-structured SVGs using a path-level latent representation; then transfer style knowledge from a customized text-to-image diffusion model into the vector model through rendered-image supervision. The key payoff is that designers could obtain a whole set of consistently styled icons or illustrations from text prompts, without per-image optimization and without the overfitting that occurs when a small style dataset is used to fine-tune a vector model directly. The paper supports the claim with quantitative comparisons and a user study across 200 trained styles.","feed_headline":"Text-to-vector generation learns a new style from just a few sample SVGs","feed_subtitle":"A two-stage pipeline separates content from style, so any text prompt yields matching vector art in seconds.","key_machinery":"The load-bearing mechanism is the path-level SVG latent representation combined with style distillation through a differentiable rasterizer. Each SVG is encoded as a sequence of path embeddings $s_0=(P_1,\\dots,P_m)$, $m\\le 32$, where each path embedding $P_i=(z_i,C_i,Tr_i)$ consists of a latent code from a pretrained path-level SVG autoencoder, a uniform color, and a transformation; a transformer-based diffusion backbone with cross-attention to CLIP text embeddings generates this sequence under the DDPM objective. In the second stage, a text-to-image diffusion model fine-tuned on style examples (via LoRA) is paired with a structure-control module that uses the Canny edge map of the rendered SVG, producing a customized image that matches the generated SVG's structure. The reparameterization trick predicts the denoised SVG tensor $\\hat{\\mathbf{s}}^g_0$, which is rendered to an image $\\hat{\\mathbf{I}}^g_0$; the image loss $\\mathcal{L}_{\\text{img}}=\\omega_t\\|\\hat{\\mathbf{I}}^g_0-\\mathbf{I}^c_0\\|^2$ together with the diffusion loss transfers the style into the vector model.","core_discovery":"The central discovery is that content and style can be disentangled in text-to-vector generation by training in two separate stages. In the first stage, a path-level T2V diffusion model is trained on black-and-white vector icons so it learns the structural regularities of clean SVG designs. In the second stage, the style of a few exemplar SVGs is captured by fine-tuning a text-to-image diffusion model, and this customized image model is used to generate many diverse styled images that become augmented training data for the vector model. The vector model is updated by an image-level loss between its rendered SVG predictions and these customized images, so the style flows from the image diffusion prior into the vector generator. After training, appending a learned style token to a text prompt produces an SVG in that style in a feed-forward manner. If correct, this makes the pipeline the first feed-forward T2V method with style customization.","pith_inferences":["The practical ceiling of the approach is set by the pretrained path-level autoencoder: because it was trained on black-and-white icons, styles that rely on gradients, strokes, or complex decorative geometry may not be representable, so extending to richer styles would likely require retraining the autoencoder on colored or decorative SVG data.","The method inherits the semantic limits of its training data and of the text-to-image teacher: prompts outside the FIGR-8 vocabulary, such as 'cello', produce inaccurate objects, so collecting a larger, better-annotated SVG dataset is the most direct route to broader prompt coverage.","A testable extension is to evaluate style fidelity on exemplars whose style lives in stroke texture or non-geometric detail; the current image-level loss and black-and-white autoencoder would likely capture silhouette and palette but lose such texture.","The same distillation scheme could be applied to other vector formats or to per-subject, not just per-style, customization, since the rendered-image bridge does not depend on SVG-specific code generation."],"forward_implications":["One trained model can serve many styles: each style is bound to a unique token, and at inference that token is simply appended to the text prompt.","Generation is feed-forward after training, taking about 25 seconds per SVG on an A6000, compared with tens of minutes for optimization-based baselines.","A new style can be added with only a few exemplars, either by full-model fine-tuning or by learning LoRA adapters that avoid introducing a new style token.","The two-stage design prevents the content collapse seen when a vector model is fine-tuned directly on a small style set, as shown by the text-alignment results in the ablation."],"supporting_citations":[{"why":"Supplies the pretrained path-level SVG autoencoder and neural path representation used to encode and decode SVG paths.","marker":"[Zhang et al. 2024]"},{"why":"Provides the DreamBooth fine-tuning recipe for associating a unique token with a user-provided style or concept.","marker":"[Ruiz et al. 2022]"},{"why":"Provides LoRA parameter-efficient fine-tuning used to create per-style adapters for both the image and vector models.","marker":"[Hu et al. 2021]"},{"why":"Provides the ControlNet structure-control module that keeps customized images aligned with the rendered SVG structure.","marker":"[Zhang et al. 2023a]"},{"why":"Provides the differentiable rasterizer that renders SVG predictions into images for the image-level loss.","marker":"[Li et al. 2020]"},{"why":"Defines the DDPM training objective and cosine noise schedule used to train the vector denoiser.","marker":"[Ho et al. 2020]"},{"why":"Supplies the FIGR-8-SVG dataset of black-and-white vector icons used in Stage 1 training.","marker":"[Clouâtre and Demers 2019]"},{"why":"Provides the reparameterization trick used to predict the denoised SVG tensor from a noisy representation.","marker":"[Song et al. 2020]"},{"why":"Provides the latent text-to-image diffusion model used as the base for style fine-tuning and as the image diffusion prior.","marker":"[Rombach et al. 2022]"}],"fun_headline_variants":["Two-stage pipeline gives text-to-vector style control from a few SVGs","Style-customized SVG generation via image priors and diffusion","Disentangling content and style for feed-forward vector art","Feed-forward style customization for text-to-vector generation","Custom vector styles from few examples, no optimization needed"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The pretrained path-level SVG compression network, trained on simple black-and-white icons, must be able to encode and decode the colored, decorative, or structurally complex paths that appear after style customization; if it cannot, the style distillation has no way to produce valid vector output.","fun_headline_variants_meta":{"raw":{"variants":["Two-stage pipeline gives text-to-vector style control from a few SVGs","Style-customized SVG generation via image priors and diffusion","Disentangling content and style for feed-forward vector art","Feed-forward style customization for text-to-vector generation","Custom vector styles from few examples, no optimization needed"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000304,"raw_usage":{"total_tokens":1775,"prompt_tokens":1002,"completion_tokens":773,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":618,"completion_tokens_details":{"reasoning_tokens":689}},"tokens_in":618,"tokens_out":773,"duration_ms":6479,"temperature":1.0,"reasoning_tokens":689,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T21:07:06.282474+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take colored, gradient-filled, or multi-layer SVG style exemplars, pass them through the path-level encoder and decoder used by the pipeline, and measure how much of the style survives reconstruction; if decoded SVGs lose color, decoration, or layer structure enough that the style becomes unrecognizable, the pretrained latent space is the bottleneck and the feed-forward customization claim fails for those styles.","supporting_citations":[],"review_version":1}