{"id":"2cb77544-cfa9-4731-b6e2-c8bd65f06147","arxiv_id":"2411.17832","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"SVGDreamer++ uses SAM-based hierarchical masks and adaptive path control to generate text-guided SVGs that are more editable and visually detailed.","lead":"SVGDreamer++ improves text-guided SVG generation by adding SAM-based hierarchical masks and adaptive path-count control to an existing diffusion-distillation pipeline. It reports better editability and visual quality over prior baselines, but the evaluation lacks error bars and released code.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Table 1's FID/PSNR reference set is undefined; if it is the LDM image used by HIVE as target, the SOTA claim measures reconstruction rather than text-to-SVG generation.","rationale":"The reader's weakest_assumption (adaptive vector primitive control) is legitimate but secondary: even if that heuristic is imperfect, the claimed SOTA could still hold. The evaluation-reference issue attacks the main evidence directly, because the headline claim is a claim about metric superiority in Table 1. If the FID/PSNR reference is the same Stable Diffusion image that HIVE is explicitly optimized against, the comparison is biased by construction, and no amount of qualitative examples fully repairs that. I therefore partially agree with the reader: the conditional verdict is appropriate, but the primary condition should be a transparent, fair evaluation protocol, not just additional ablations of the primitive-control module. The paper does provide plausible qualitative evidence and promises code release, which counts in its favor; however, the quantitative SOTA assertion cannot be accepted without knowing what the metrics compare against.","tokens_in":27842,"tokens_out":6430,"duration_ms":65507,"concrete_test":"Obtain the exact prompt set and the reference images used for FID/PSNR from the authors. Then recompute Table 1 under two conditions: (a) with a reference distribution generated by Stable Diffusion using seeds held out from the optimization; (b) adding a strong raster-to-vector baseline (e.g., LIVE) that is given the same LDM output as input rather than optimizing from random primitives. If SVGDreamer++'s margin collapses in either condition, the 'best across all metrics' claim is an artifact of the evaluation reference; if the margin persists, the concern is resolved.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The paper's central claim is 'best performance across all evaluated metrics' (Table 1). FID and PSNR are reference-based metrics, yet the full text never states which images form the FID reference set, nor what the PSNR reference image is. The pipeline itself generates a raster I from Stable Diffusion and optimizes the SVG to match I through HIVE (Eq. 8). If the evaluation reference is I, or any set of SD outputs on the same prompts, then SVGDreamer++ is rewarded for copying its own supervision target, while methods that do not receive that reference (CLIPDraw, Evolution, VectorFusion) are structurally disadvantaged. The reported PSNR values (e.g., 15.80 vs. 8.35) would then reflect reconstruction fidelity to a method-specific raster, not general text-to-SVG quality. This is not merely a missing detail: it is the evidence for the headline claim, and Table 1 does not isolate the contributions of HIVE or adaptive control through quantitative ablations. The concern is correctable by a protocol disclosure, but it is load-bearing because the SOTA assertion depends entirely on these numbers.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents SVGDreamer++, a text-guided SVG generation method that extends the authors' earlier SVGDreamer with two new components: a Hierarchical Image VEctorization (HIVE) strategy that uses Grounded-SAM and SAM masks to supervise vectorization at both object and part levels, and an Adaptive Vector Primitives Control algorithm that dynamically splits, clones, or prunes vector paths during optimization based on gradient magnitudes. The method combines these components with the previously proposed Vectorized Particle-based Score Distillation (VPSD) and Reward Feedback Learning to improve visual quality, diversity, and editability. The authors claim state-of-the-art performance on six metrics, demonstrate object- and part-level editing, and show applications in poster design and vector asset generation.","tokens_in":28088,"tokens_out":2822,"duration_ms":27028,"significance":"If the central claims hold, the paper makes a useful contribution to text-to-SVG generation by addressing two practically important shortcomings: the lack of editability in the output SVGs and the difficulty of setting the number of vector primitives in advance. The HIVE framework's use of segmentation priors for part-level decomposition is a plausible and well-motivated idea, and the adaptive primitive control is a novel mechanism for controlling SVG complexity during optimization. The paper is also commendable for including qualitative comparisons of three vectorization approaches, ablations of particle counts, and a stated intention to release code. However, the main quantitative claim is currently supported by an incompletely specified evaluation protocol, and the contributions of the two new modules are not isolated by quantitative ablations; these issues need to be resolved before the claims can be fully accepted.","major_comments":[{"comment":"The reference set for FID and the reference image for PSNR are never defined. The paper states that HIVE optimizes the SVG to match a raster image I generated by a latent diffusion model (Eq. 8), and the evaluation section (Section 6, 'Evaluation Metrics') only names the metrics without specifying the reference distribution or reference image. If the FID reference set and the PSNR reference are the LDM-generated images used internally by each method, then methods that receive a direct raster target (VectorFusion, SVGDreamer, SVGDreamer++) are structurally favored, while CLIP-based methods that do not use such a target are disadvantaged. This is load-bearing because the headline claim of 'best performance across all evaluated metrics' rests entirely on Table 1. Please specify the reference set, report the number of prompts and seeds, and provide the protocol so that the comparison is method-independent.","section":"Section 6, Table 1"},{"comment":"The adaptive vector primitive control is central to the claimed visual quality improvement, but its effectiveness is supported only by qualitative visualizations (Figs. 11, 12). There is no quantitative ablation in Table 1 or elsewhere that isolates the contribution of this module, and no sensitivity analysis for the hand-tuned thresholds tau_opacity, tau_c, and tau_a. Since the algorithm's decision rule assumes that high gradient magnitude in the HIVE loss reliably indicates regions needing more paths, an ablation that compares the full method with and without adaptive control, along with a sensitivity study for the thresholds, is necessary to validate this load-bearing assumption.","section":"Section 4.2, Algorithm 1"},{"comment":"Table 2 reports iteration counts and wall-clock times for VPSD with and without Reward Feedback Learning, but it does not report any quality metric for these two configurations. The text states that ReFL 'improves the aesthetic score of the SVG' and refers to Table 1, but Table 1 compares full methods rather than a controlled with/without-ReFL ablation. Please provide a quantitative ablation (e.g., CLIP score, aesthetic score, or HPS) for the same prompt set with and without ReFL, otherwise the aesthetic-improvement claim is not supported by the presented evidence.","section":"Section 6.2.5, Table 2"},{"comment":"The paper describes three prompt sets (single object, single object with surroundings, multiple objects) but does not state how many prompts are in each set, how many optimization runs per prompt are used, or whether Table 1 aggregates over all three sets. No error bars or variances are reported for any of the six metrics. Given that the reported differences between some methods are moderate (e.g., CLIP score 0.3001 vs. 0.3093), the absence of any statistical variability measure makes it difficult to assess whether the differences are meaningful. Please provide the number of prompts, seeds, and per-metric variances or confidence intervals.","section":"Section 6, Dataset and Evaluation Metrics"}],"minor_comments":[{"comment":"There are several typos and inconsistencies that should be corrected: 'SVGDremaer++' in Figure 1, 'purning' in Algorithm 1, 'editied' in Section 6.1.3, 'VPSD v.s. LSDS v.s. ASDS' in Section 6.2.3, and 'Low-ploy' in Figure 15.","section":"Throughout"},{"comment":"The definition of the foreground attention maps M_FG is unclear: the softmax is written as softmax(Q K_i^T)/sqrt(d), but it is not specified over which dimension the softmax is applied, and the threshold used later to convert attention maps into binary masks is not stated.","section":"Section 3.1.1, Eq. 1"},{"comment":"The reward feedback loss in Eq. 6 is written as an expectation over y of psi(r(y, g_phi_est(y))), but the text describes a reweighting of samples collected from the LoRA model. The connection between the expectation form and the 'reweighted samples' description should be made explicit, and the role of the ReLU map psi should be clarified.","section":"Section 3.2, Eq. 6"},{"comment":"The operations RemovePath, SplitPath, and ClonePath are not formally defined. In particular, it is not specified how a path is split into two paths (whether control points are duplicated, whether colors are inherited, and how the new paths are initialized) or how a cloned path is placed relative to the original.","section":"Algorithm 1"},{"comment":"The text says 'we limit our testing to a maximum of 6 particles,' but Figure 13 shows results for 1, 4, 8, and 16 particles, which exceeds this maximum. This inconsistency should be resolved, either by correcting the text or by explaining how the 16-particle results were produced.","section":"Section 6.2.4, Figure 13"},{"comment":"The poster-design comparison with raster T2I methods is presented only qualitatively. Since the paper emphasizes the advantages of vector posters in text clarity and editability, a small quantitative or user-study evaluation would strengthen this section, but this is not essential to the main claims.","section":"Section 6.3.1, Figure 16"}],"recommendation":"major_revision","confidential_remarks":"The paper is an extension of the authors' CVPR 2024 SVGDreamer paper. The incremental novelty over the conference version is primarily HIVE and the adaptive vector primitive control, which are the two components whose evaluation needs strengthening. The evaluation protocol issue for Table 1 is the most serious concern: if the reference set is method-specific LDM outputs, the headline comparison is not a fair text-to-SVG comparison. This is fixable by reporting a clear protocol, but it is load-bearing. Please ensure that the revision includes quantitative ablations for the two new modules and a sensitivity analysis for the adaptive control thresholds."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"What you should know: the new pieces are real, but the headline numbers are not yet trustworthy. HIVE (SAM-based hierarchical vectorization) and Adaptive Vector Primitives Control are legitimate technical additions, and the paper shows compelling editing demos. But Table 1's FID/PSNR numbers are uninterpretable because the reference set is never specified, and the pipeline is explicitly supervised by a Stable Diffusion raster. The stress-test concern holds up.\n\nWhat the paper does well: the HIVE idea of using Grounded-SAM for object-level masks and SAM for part-level masks is a sensible upgrade over the attention-only masks in the authors' prior SVGDreamer. The adaptive path splitting/cloning heuristic is novel and appears to work in the qualitative figures. The writing is clear, the limitation statement is honest, and the claim that editability is limited by the T2I model is a fair caveat.\n\nWhere the soft spots are: first, the evaluation. The paper never says what images form the FID reference set or what the PSNR reference is. If it is the SD-raster used by HIVE as supervision, then SVGDreamer++ is rewarded for reconstructing a target that CLIPDraw and Evolution never see. That is not a missing detail; it is the foundation of the SOTA claim. Second, there are no error bars, no sensitivity analysis on the three hand-tuned thresholds, and no quantitative comparison with four concurrent methods cited in the related work. Third, the ablation study is qualitative; Figure 10-12 show behavior but do not isolate how much HIVE versus adaptive control contributes to the Table 1 numbers. These are correctable, but they are not cosmetic. The central idea is plausible, but the evidence as presented is not enough to accept the superiority claim.\n\nWho gets value: practitioners in text-to-SVG and generative vector design. The techniques are worth knowing about, but treat the quantitative comparison as provisional until the protocol is disclosed.\n\nRecommendation: this deserves a serious referee. The novelty is sufficient, but the referee should demand the evaluation protocol, error bars, and ablations that separate the new components. If those are supplied, the paper could be solid; as it stands, the SOTA claim is not supported.","headline":"SVGDreamer++ adds two genuinely new components and a coherent pipeline, but its state-of-the-art claim rests on an evaluation protocol that is never defined and may reward copying the supervision image rather than text-to-SVG generation.","tokens_in":28613,"tokens_out":2261,"would_cite":false,"duration_ms":23046,"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":"Hierarchical masks make generated SVGs editable and diverse.","keywords":["vector graphics","SVG generation","text-to-SVG","hierarchical vectorization","segmentation-guided optimization","score distillation","editable vector graphics","diverse generation"],"falsifier":"Run the full SVGDreamer++ pipeline on a prompt set with fine-grained texture such as fur, hair, or foliage, and compare the variant with adaptive primitive control disabled against the variant with it enabled while holding all other hyperparameters fixed; also vary the gradient threshold $\\tau_c$ across a range. If FID and PSNR do not improve on the high-texture set, or if the threshold sensitivity is so sharp that small changes reverse the gain, the adaptive-control claim is not robust. A simpler specific check is to render the gradient map of the HIVE loss for a target whose geometry is already perfectly covered but whose color is wrong; if the algorithm still splits paths, the heuristic is misfiring.","tokens_in":27635,"feed_emoji":"🎨","tokens_out":6035,"duration_ms":50552,"temperature":0.7,"pith_summary":"The paper aims to show that text-guided SVG generation can produce outputs that are not only visually competitive with raster text-to-image results but also genuinely editable, meaning a designer can select, recolor, recombine, or replace individual objects and object parts. To do this, it introduces a hierarchical vectorization step (HIVE) that uses image-segmentation masks to supervise separate groups of Bézier paths, and a particle-based score distillation (VPSD) that treats several sets of SVG parameters as samples from a distribution. It also changes the number of paths during optimization, splitting or cloning paths in regions where the loss gradient suggests under- or over-representation. The paper reports that the combined method outperforms prior text-to-SVG systems on all six evaluated metrics and supports six vector styles.","feed_headline":"Hierarchical masks make text-to-SVG output editable and diverse","feed_subtitle":"Segmentation-guided vectorization plus particle-based distillation beats prior text-to-SVG methods on six metrics.","key_machinery":"The load-bearing mechanism is the hierarchical mask loss $L_{HIVE} = \\sum_i \\|\\hat{M}_i \\odot I - \\hat{M}_i \\odot x\\|_2^2 + \\sum_i\\sum_j \\|\\tilde{M}^j_i \\odot I_i - \\tilde{M}^j_i \\odot x_i\\|_2^2$, where $\\hat{M}_i$ are object masks and $\\tilde{M}^j_i$ are part masks. The masks come from a two-stage process: coarse masks from a text-grounded segmentation, then fine masks conditioned on the current control-point coordinates. Because the loss is masked per region, each path group is supervised only inside its own semantic region, which is what enforces decoupling and editability. The adaptive primitive control (Algorithm 1) is the companion mechanism: it prunes near-transparent paths, splits over-represented large paths, and clones under-represented small paths, using the gradient magnitude of the HIVE loss to locate regions that need more primitives.","core_discovery":"SVGDreamer++ proposes that the right supervision signal for text-to-SVG is not the full raster rendering but a hierarchy of masks: object-level masks obtained by grounding each noun in the prompt, and part-level masks obtained by conditioning a segmentation model on the current control-point positions of each object. With these masks as per-region losses, vector paths stay inside their semantic region, so the final SVG can be edited at object and part level. The same optimization is driven by VPSD, which maintains k particles (sets of control points and colors), estimates their distribution with a LoRA-adapted diffusion model, and reweights particles by a pretrained aesthetic reward, yielding diverse, less saturated images with faster convergence. Combined with adaptive primitive control, which splits large paths that cover too much geometry and clones small paths that cover too little based on the gradient magnitude of the HIVE loss, the result is a text-to-SVG pipeline that the authors claim achieves the best scores on FID, PSNR, CLIP, BLIP, Aesthetic, and HPS among the methods they compare.","pith_inferences":["The gradient-magnitude heuristic for path splitting and cloning is really a proxy for reconstruction error; one could test it directly by comparing against a policy that adds paths based on masked reconstruction error or segmentation uncertainty, to see whether gradient magnitude is the right trigger.","The hierarchy idea could be pushed one level further: if masks are available, the same per-region loss could be used to learn a reusable compositional prior, such as a library of independently editable object prototypes, rather than optimizing each prompt from scratch.","Because the method's editability ceiling is set by the text-to-image model's ability to separate concepts in its attention, improvements in text-to-image grounding should transfer automatically; the paper itself notes this dependence in its limitations.","A natural stress test is prompts with several overlapping same-category objects, such as two astronauts, where attention-based initialization and part masks may conflict; the paper shows one such case, but systematic scaling of the number of objects would reveal where the decomposition degrades."],"forward_implications":["If HIVE works as claimed, vector objects generated from multi-noun prompts can be decoupled into independently editable components, enabling recombination of foregrounds and backgrounds from different generated SVGs.","Part-level masks would allow local edits, such as swapping a character's cloak or weapon, without re-running generation or manually tracing paths.","VPSD's particle distribution would yield multiple SVG samples for one prompt with more diversity and less color over-saturation than SDS-based methods, at roughly half the optimization time when reward feedback is used.","Adaptive primitive control would remove the need to choose a fixed path count in advance, improving detail where needed while keeping optimization faster than a uniformly large path budget.","With the six primitive-style settings (icon, pixel art, low-poly, painting, sketch, ink-wash), the same pipeline can serve as a vector-asset generator for posters, stickers, and item art."],"supporting_citations":[{"why":"Supplies the differentiable rasterizer that lets SVG path and color parameters be optimized by backpropagation.","marker":"[13]"},{"why":"Defines the text-to-SVG baseline with SDS and LIVE-based vectorization that VPSD and HIVE are designed to improve.","marker":"[7]"},{"why":"Provides the attention-map-based primitive initialization strategy that HIVE reuses.","marker":"[8]"},{"why":"Is the conference version of SVGDreamer whose SIVE and VPSD components this paper extends.","marker":"[9]"},{"why":"Introduces the SDS loss that VPSD replaces as the core distillation mechanism.","marker":"[20]"},{"why":"Presents the variational score distillation and particle-based distribution modeling that VPSD adapts to vector graphics.","marker":"[56]"},{"why":"Supplies the part-level segmentation masks used for fine-grained vectorization supervision.","marker":"[24]"},{"why":"Supplies object-level masks grounded in text nouns for coarse vectorization supervision.","marker":"[59]"},{"why":"Provides the pretrained reward model used to reweight vector particles and improve aesthetics.","marker":"[23]"},{"why":"Provides the low-rank adaptation network used to estimate the distribution of SVG parameters in VPSD.","marker":"[22]"}],"fun_headline_variants":["Hierarchical masks and VPSD distillation boost SVG editability and diversity","SVGDreamer++: hierarchical masks enable editing, VPSD adds diversity","Object-level masks and particle distillation improve text-to-SVG quality","HIVE masks and VPSD distillation make SVG generation editable and diverse","Segmentation-guided vectorization and particle distillation boost SVG diversity"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method assumes that a large gradient magnitude in the HIVE loss marks a spot where a path should be split or cloned, rather than a spot that needs a different color, opacity, or a corrected boundary; if that mapping is unreliable, added paths land in the wrong places and the reported quality gain from adaptive primitive control could disappear.","fun_headline_variants_meta":{"raw":{"variants":["Hierarchical masks and VPSD distillation boost SVG editability and diversity","SVGDreamer++: hierarchical masks enable editing, VPSD adds diversity","Object-level masks and particle distillation improve text-to-SVG quality","HIVE masks and VPSD distillation make SVG generation editable and diverse","Segmentation-guided vectorization and particle distillation boost SVG diversity"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000748,"raw_usage":{"total_tokens":3400,"prompt_tokens":1079,"completion_tokens":2321,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":695,"completion_tokens_details":{"reasoning_tokens":2228}},"tokens_in":695,"tokens_out":2321,"duration_ms":14602,"temperature":1.0,"reasoning_tokens":2228,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T11:46:52.801776+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the full SVGDreamer++ pipeline on a prompt set with fine-grained texture such as fur, hair, or foliage, and compare the variant with adaptive primitive control disabled against the variant with it enabled while holding all other hyperparameters fixed; also vary the gradient threshold $\\tau_c$ across a range. If FID and PSNR do not improve on the high-texture set, or if the threshold sensitivity is so sharp that small changes reverse the gain, the adaptive-control claim is not robust. A simpler specific check is to render the gradient map of the HIVE loss for a target whose geometry is already perfectly covered but whose color is wrong; if the algorithm still splits paths, the heuristic is misfiring.","supporting_citations":[{"cited_title":"Differentiable vector graphics rasterization for editing and learning,","cited_arxiv_id":null,"evidence_quote":"Supplies the differentiable rasterizer that lets SVG path and color parameters be optimized by backpropagation."},{"cited_title":"Vectorfusion: Text- to-svg by abstracting pixel-based diffusion models,","cited_arxiv_id":null,"evidence_quote":"Defines the text-to-SVG baseline with SDS and LIVE-based vectorization that VPSD and HIVE are designed to improve."},{"cited_title":"Diffsketcher: Text guided vector sketch syn- thesis through latent diffusion models,","cited_arxiv_id":null,"evidence_quote":"Provides the attention-map-based primitive initialization strategy that HIVE reuses."},{"cited_title":"Svgdreamer: Text guided svg generation with diffu- sion model,","cited_arxiv_id":null,"evidence_quote":"Is the conference version of SVGDreamer whose SIVE and VPSD components this paper extends."},{"cited_title":"Dreamfusion: Text-to-3d using 2d diffusion,","cited_arxiv_id":null,"evidence_quote":"Introduces the SDS loss that VPSD replaces as the core distillation mechanism."},{"cited_title":"Segment anything,","cited_arxiv_id":null,"evidence_quote":"Supplies the part-level segmentation masks used for fine-grained vectorization supervision."},{"cited_title":"Grounded sam: Assembling open-world models for diverse visual tasks,","cited_arxiv_id":null,"evidence_quote":"Supplies object-level masks grounded in text nouns for coarse vectorization supervision."},{"cited_title":"Imagereward: Learning and evaluating human preferences for text-to-image generation,","cited_arxiv_id":null,"evidence_quote":"Provides the pretrained reward model used to reweight vector particles and improve aesthetics."},{"cited_title":"LoRA: Low-rank adaptation of large language models,","cited_arxiv_id":null,"evidence_quote":"Provides the low-rank adaptation network used to estimate the distribution of SVG parameters in VPSD."}],"review_version":1}