SVGEval: A Vision-Grounded Framework for Perceptual-Quality Benchmarking and Evaluation in Text-to-SVG Generation¶

Conference: ECCV 2026
Paper: ECCV 2026
Area: LLM Evaluation / Multimodal VLM
Keywords: SVG Generation Evaluation / Multimodal Large Models / Vision Grounding / Explainable Scorer / Benchmark Construction

TL;DR¶

SVGEval moves the evaluation target of text-to-SVG from the SVG code to the rendered outcome that users actually see: it builds a 500-instance binary-diagnosis plus 500-instance 1–5 scoring benchmark over a four-dimension rubric (Aesthetic, Semantic, Spatial, Structural) with multi-round human annotation and expert adjudication, then distills an explainable scorer that reads both the rendering and the code and outputs dimension-wise scores with evidence-grounded rationales — cutting MAE from 0.71 (best general model) to 0.58 and lifting ±1 adjacent accuracy from 66.6% to 72.0%.

Background & Motivation¶

Scalable Vector Graphics (SVG) encodes visual content as a structured program of paths, primitives, and hierarchical groups, which makes it resolution-independent, compact to store, and highly editable; that is why it is a fundamental format for icons, diagrams, UI design, and infographic authoring, where geometric fidelity and reusability matter most. Multimodal large models have recently pushed text-to-SVG generation forward rapidly — StarVector, OmniSVG, LLM4SVG, and SVGThinker can already produce complex vector graphics from natural-language instructions. Generation capability, however, has outrun evaluation: unlike text-to-image, where widely adopted quantitative metrics and protocols have been studied extensively, SVG generation still lacks a unified and reliable standard, so comparisons across methods are inconsistent and the "improvements" reported under one evaluation setting may not translate into human-perceived quality.

The dominant practice today is to rasterize the SVG into a bitmap and then apply image-generation metrics — MSE, LPIPS, FID, CLIPScore — to quantify fidelity and prompt alignment. This raster-centric paradigm is fundamentally misaligned with the nature of SVG. SVG quality is largely determined by structured geometry and spatial composition rather than local texture, and the perceptual differences that actually matter for vector graphics — slight geometric distortions, misaligned spacing, broken symmetries, inconsistent path structures — are weakly reflected or simply ignored by pixel-level errors and deep feature similarities. Conversely, rasterization introduces arbitrary factors (resolution, anti-aliasing, stroke rendering, sampling) that can dominate MSE/LPIPS/FID without corresponding to any meaningful change in the underlying vector structure. CLIP-based similarity fails in the other direction: it mainly captures coarse image-level semantic alignment and is insensitive to the fine-grained layout rationality and geometric correctness that drive the usability and perceived quality of SVG. Existing SVG benchmarks, meanwhile, are mostly task-driven (understanding, editing, generation), telling you whether a task was completed rather than diagnosing what is badly drawn.

This paper's angle is to change the anchor of judgment: since SVG is ultimately consumed visually by humans, evaluation should treat the rendered outcome as the primary target while keeping the structural cues in the vector code. Core idea: define a four-aspect rubric (Aesthetic, Semantic, Spatial, Structural) disambiguated by a hierarchical attribution rule — explicit errors are charged to Structural/Spatial first and Aesthetic is defined only as a residual; let the evaluator see the prompt, the rendering, and the SVG code together while all gold labels are defined over the rendered outcome; then use this benchmark both to systematically audit current MLLMs and to train an explainable scorer via reasoning distillation that reconstructs evidence from a given human score.

Method¶

Overall Architecture¶

SVGEval interlocks three parts: a four-aspect evaluation protocol with a hierarchical attribution principle, a difficulty-progressive two-part benchmark (Part I: 500 binary-diagnosis instances; Part II: 500 quality-assessment instances scored 1–5), and an explainable scorer built on the same data. The construction chain runs as follows: starting from a large corpus of real SVG–text pairs (StarVector and SVG Repo), an LLM extracts topic descriptors and clusters the corpus to obtain an empirical topic distribution; stratified sampling then follows that distribution to collect 1,000 prompts spanning simple icons to complex diagrams. These prompts are fed to several LLMs/MLLMs of varying ability, and the generations are mixed with real SVGs from the corpus into a candidate pool of roughly 10k samples that forms a continuous quality gradient from severely broken to near-correct. The pool is then curated and annotated under the two task formats. At evaluation time, the model sees the prompt, the rendered image, and the SVG source code simultaneously, while every gold label is defined with respect to the rendering.

%%{init: {'flowchart': {'rankSpacing': 24, 'nodeSpacing': 28, 'padding': 6, 'wrappingWidth': 400}}}%%
flowchart TD
    A["Four-aspect rubric + hierarchical attribution<br/>four complementary dimensions, explicit errors first"] --> B["Rendering as the primary evaluation target<br/>prompt + rendering + SVG code"]
    B --> C["Difficulty-progressive two-part benchmark<br/>Part I binary diagnosis + Part II 1-5 scoring"]
    C --> D["Score-conditioned reasoning distillation<br/>teacher reconstructs evidence from the human gold score"]
    D --> E["Explainable scorer<br/>dimension-wise scores + rationales"]
    E --> F["Benchmark and scorer evaluation<br/>Accuracy / MAE / Adj. Acc."]

Key Designs¶

1. Four-aspect rubric with hierarchical attribution: splitting "how good is it" into attributable aspects

A single "quality" score cannot answer where a sample fails, so SVGEval defines four complementary dimensions: Aesthetic (visual clarity and overall presentation quality — color harmony, composition, style coherence, perceptual refinement; it targets "how good it looks" rather than strict correctness), Semantic (conceptual alignment with the text prompt only, independent of how well it is drawn), Spatial (2D layout correctness on the canvas — the accuracy of each element's absolute position, the accuracy of relative positions among elements, and the overall spatial harmony of the global layout), and Structural (geometric integrity and construction precision — closed paths, continuity, shape completeness, engineering-level correctness of geometry, and the balance between complexity and neatness). The dimensions are deliberately separable: a geometrically perfect shape can still be semantically wrong (high Structural, low Semantic), and two individually correct objects can be spatially misplaced (high Structural, low Spatial), so failures can be attributed to specific perceptual aspects. Notably, vector-legality issues such as unclosed paths or anomalous coordinates are not given a separate "code validity" dimension — they are folded into Structural, because what the paper cares about is whether a legality problem produces a visible geometric defect after rendering.

Independence does not come for free at the perceptual level: broken shapes and chaotic layouts also degrade overall visual appeal, so Aesthetic easily degenerates into a catch-all bucket. The benchmark therefore adopts a hierarchical attribution principle — explicit errors first: if a negative visual impression is clearly caused by geometric defects (gaps, broken strokes, unclosed paths) or spatial defects (unintended overlaps, off-canvas placement), the penalty is charged to Structural or Spatial first; Aesthetic is defined as a residual, reserved for properties not captured by structural/spatial correctness, such as style maturity, a pleasing palette, smoothness and cleanliness, and overall finish. The same principle is written into the distillation prompts of the scorer, so the teacher's rationales follow exactly the annotation logic of the benchmark.

2. Rendering as the primary evaluation target with a three-way input: labels look at the picture, judgment can consult the code

"Vision-grounded" has two halves here, and conflating them leads to a misreading — that SVGEval simply rasterizes the SVG and applies image metrics, which is precisely what it argues against. The first half is the label anchor: although SVG is represented as a program, perceived generation quality is ultimately determined by the rendered outcome, so all criteria are defined over what humans see in the final image rather than over SVG syntax validity or code elegance, and all four dimension scores are assigned by humans looking at the rendering. The second half is input visibility: in real workflows the model often has access to both the rendered image and the underlying structured representation, so SVGEval lets the evaluator see (i) the prompt, (ii) the rasterized rendering, and (iii) the SVG source code, allowing perceptual evidence (broken strokes, unintended overlaps) to cross-validate code-level cues (unclosed paths, anomalous coordinates, layer ordering). This stands in direct opposition to code-centric SVG evaluation. The asymmetry is empirically backed: removing the rendering and keeping only the code (noimg) degrades MAE from 0.5772 to 0.9557 and Adj. Acc. from 71.95% to 57.75%, collapsing the model into format imitation with unstable calibration; removing only the code (nosvg) costs just 0.045 MAE overall but hurts Spatial and Structural much more (Spatial 0.5167 → 0.6400). The rendering is thus a prerequisite for perceptual judgment, while the code contributes complementary cues on the SVG-specific dimensions.

3. A difficulty-progressive two-part benchmark: separate output formats for "can see it" and "can rate it well"

The benchmark is deliberately split into two parts of 500 instances each, because detecting a defect and producing a calibrated ordinal score are different abilities that a single question format cannot separate. Part I is perception / Yes–No: binary diagnostic questions probe basic perceptual properties; a multimodal model first proposes candidate questions and preliminary answers given the prompt, rendering, and code, and human annotators then revise and validate both the questions and the Yes/No labels for correctness and clarity. To remove the response bias of LLMs (a tendency toward either Yes or No), the dataset is explicitly curated to keep an approximately balanced positive/negative ratio (close to 1:1) across question types. Part II is quality assessment / scoring: each of the four dimensions is rated on a 1–5 Likert scale; to keep the benchmark discriminative across quality regimes, the sampling ratio between real and model-generated SVGs is set to 1:4 so that scores are relatively balanced across score bins for every dimension. Each instance is independently scored by multiple annotators on all dimensions, and disagreements are resolved by expert adjudication to produce the final gold scores. The two parts serve two stated purposes: balanced supervision signals offset model-specific response bias, and high-quality human annotation supports fine-grained diagnosis.

The two parts also use different metrics, both designed for ordinal scales. Part I is binary classification and reports Accuracy (overall and optionally per question type). Part II predicts ordinal scores \(S_{\text{pred}} \in \{1,\dots,5\}\), and because minor disagreements are expected in subjective ratings it reports two complementary metrics — mean absolute error for the average deviation magnitude, and adjacent accuracy, which tolerates minor subjective variation by counting predictions within ±1 as acceptable:

\[\text{MAE}=\frac{1}{N}\sum_{i=1}^{N}\left|S_{\text{pred}}^{(i)}-S_{\text{gt}}^{(i)}\right| \qquad \text{Adj. Acc.}=\frac{1}{N}\sum_{i=1}^{N}\mathbb{1}\!\left(\left|S_{\text{pred}}^{(i)}-S_{\text{gt}}^{(i)}\right|\le 1\right)\]

⚠️ Both equations were garbled by the text extraction of the cached full paper; they are reconstructed from the stated definitions ("a prediction is counted as correct if it falls within ±1 of the human gold score") — refer to the original paper for notational details. Main results report both overall performance and a per-dimension breakdown, which is what provides the diagnostic granularity.

4. Score-conditioned reasoning distillation: the teacher reconstructs evidence from the human score before it is distilled into the student

Human annotation naturally provides gold scores but rarely includes detailed, structured rationales, and a scorer that emits a number without evidence is neither auditable nor diagnosable — experiments confirm that score-only training yields limited gains and leaves the model as a black-box regressor with weak diagnostic capability. SVGEval's answer is score-conditioned rationalization: given the prompt, rendered image, SVG code, a target dimension, and its human gold score, a teacher model (Qwen3-VL-235B-A22B) acts as an analyst under the instruction "the score on dimension \(d\) is \(s\) (assigned by human experts); identify concrete visual/code evidence that supports this score according to the rubric," producing image–rationale–score triples. Conditioning on a given score rather than letting the teacher score freely anchors the reasoning to human judgment instead of injecting the teacher's own scoring bias. To reduce cross-dimension interference — for instance, complaining about appearance while writing a Semantic rationale — the synthesis prompts impose dimension constraints: for Semantic, the teacher must ignore visual refinement (Aesthetic) and geometry/layout defects (Structural/Spatial) and focus only on concept alignment with the prompt. Rationales must also follow the benchmark's hierarchical attribution principle, keeping synthesized rationales and benchmark annotation on the same diagnostic logic. The student is supervised-fine-tuned on the Qwen3-VL 30B A3B backbone with a loss covering both rationale text generation and score tokens, and at inference time it takes only (prompt, rendering, code, dimension) and emits the dimension-wise score together with a rationale.

A Worked Example¶

Take one concrete instance from Table 4 of the paper and walk it through distillation and inference. The input is a prompt asking for "a black Bluetooth symbol: a stylized B formed by overlapping diagonal lines, thick solid black strokes," plus a rendered image and its SVG code; the target dimension is Spatial and the human gold score is 5/5. In the distillation stage the teacher receives all four items and must find evidence under the constraint that the score is already known to be 5, so it writes things like "the Bluetooth symbol… excellent spatial coherence… the two diagonal lines forming the B… match what is expected for a well-designed Bluetooth icon." Note that it searches for evidence supporting a score of 5 — the relative positions of the diagonals, the symmetry of the symbol — rather than praising the picture in general. Once this triple enters the SFT data, the mapping the student learns shifts from "image → score" to "image + code + dimension → score + the evidence backing it." At inference the student sees only (prompt, rendering, code, "Spatial") and outputs something like "Score: 5" followed by a rationale. The same table shows two other characteristic outputs: an image whose semantics are roughly right but where "a vague cup-like shape is visible… the coffee cup is not clearly depicted" receives Semantic 3/5, so the mid-level score is backed by evidence rather than assigned arbitrarily; and a cityscape scored Structural 4/5 with the rationale "perspective lines converge neatly toward a central vanishing point… slightly coarse shadows reduce structural precision, but overall execution remains high-quality," demonstrating the hierarchical-attribution convention in which a minor structural defect costs points without being fatal.

Loss & Training¶

The scorer undergoes standard supervised fine-tuning on the synthesized data; the objective minimizes the generation loss of both the rationale text and the corresponding score tokens, encouraging faithful, rubric-aligned explanations alongside calibrated scores. The teacher side uses Qwen3-VL-235B-A22B and the student backbone is Qwen3-VL 30B A3B — a choice driven by the benchmark itself: after running the full SVGEval evaluation, the strongest open-source model is selected as the base, closing the loop in which the benchmark guides model selection and training then feeds back into evaluation capability. ⚠️ The cached full text does not report learning rate, training epochs, batch size, or the training-set size — refer to the original paper / appendix.

Key Experimental Results¶

Main Results¶

The benchmark evaluates 12 representative general-purpose multimodal models plus the trained scorer. Part I reports binary accuracy; Part II reports MAE and ±1 adjacent accuracy. SVGEval-Scorer is designed for the multi-aspect scoring task, so binary diagnosis is marked "—".

Model	Binary Diagnosis Acc. (%) ↑	Scoring MAE ↓	Scoring Adj. Acc. (%) ↑
Gemini-3 Flash	89.40	0.7720	62.80
Gemini-3 Pro	88.00	0.7068	66.57
Claude Sonnet 4.5	83.80	0.8980	57.90
Doubao-Seed 1.6 Vision	82.20	1.2120	46.20
GPT-5	80.76	0.9137	57.13
Qwen3-VL-235B	78.40	0.9860	53.90
InternVL3.5	77.40	1.1200	50.50
GPT-5.2	76.80	0.9520	55.50
GLM-4.5V	75.95	1.2740	44.79
Grok-4.1-Fast	74.55	1.3614	42.47
Gemini-2.0 Flash	69.94	0.9640	54.90
GPT-4o-mini	69.20	0.9060	56.80
SVGEval-Scorer (ours)	–	0.5772	71.95

The per-dimension breakdown exposes which two dimensions carry the weakness of general models (Table 3; lower MAE is better):

Model	Aesthetic ↓	Semantic ↓	Spatial ↓	Structural ↓
Gemini-3 Pro	0.7623	0.6160	0.8400	0.6111
Gemini-3 Flash	0.6210	0.7760	0.9120	0.7778
GPT-5	0.7661	0.7840	1.2258	0.8800
Claude Sonnet 4.5	0.7903	0.8080	1.0960	0.8968
Qwen3-VL-235B	0.9758	0.7920	1.1520	1.0238
GLM-4.5V	1.6967	0.8017	1.4628	1.1360
Grok-4.1-Fast	1.7419	0.8720	1.4715	1.3651
InternVL3.5	0.9919	0.7840	1.6080	1.0952
GPT-4o-mini	0.8790	0.7280	0.9600	1.0556
SVGEval-Scorer (ours)	0.5935	0.6129	0.5167	0.5840

Ablation Study¶

Three ablations remove rationale supervision, the SVG-code input, and the rendering input in turn, validating the contribution of each ingredient:

Config	MAE ↓	Adj. Acc. (%) ↑	Spatial MAE ↓	Structural MAE ↓	Note
Full scorer	0.5772	71.95	0.5167	0.5840	vision + code + rationale supervision
w/o reasoning (nothink)	0.8330	60.30	0.8750	0.8790	dropping the rationale objective costs 0.256 MAE / 11.7 points and destroys auditability
w/o SVG input (nosvg)	0.6220	70.40	0.6400	0.6349	only 0.045 MAE overall, but Spatial / Structural suffer noticeably more
w/o image input (noimg)	0.9557	57.75	0.9187	0.9841	code only; costs 0.379 MAE / 14.2 points, the most severe degradation

Key Findings¶

"Seeing a defect" and "rating it well" are different abilities. The 12 general models reach 69.2–89.4% on Part I, which already looks decent, yet Part II MAE spans 0.71–1.36 with Adj. Acc. of only 42.5–66.6%; the two are not perfectly correlated. The clearest evidence is Gemini-3 Flash versus Gemini-3 Pro: Flash has the highest binary accuracy (89.40% vs 88.00%) but a clearly worse scoring MAE (0.7720 vs 0.7068). Detecting discrete defects does not translate into calibrated ordinal ratings.
The two SVG-specific dimensions are a systematic weakness. Almost every model has substantially larger errors on Spatial and Structural than on Aesthetic and Semantic, with Spatial most often the hardest dimension, reflecting persistent difficulty in layout/topology reasoning. The relatively stronger Semantic and Aesthetic dimensions happen to be the primary focus of most raster-centric multimodal benchmarks, suggesting those perceptual skills are better studied and steadily improving — which is exactly why raster metrics under-report the failure modes that matter most for SVG usability.
The difficulty ordering is not uniform across models. Gemini-3 Pro's Structural MAE (0.6111) is actually its best dimension, better than its Aesthetic (0.7623), whereas GPT-5's Spatial MAE (1.2258) is far worse than its Structural (0.8800). Spatial/Structural are harder in aggregate, but individual models deviate, and looking only at overall MAE hides this structural difference.
Missing-modality damage is asymmetric and localizes to dimensions. noimg causes the largest degradation (MAE +0.379, Adj. Acc. −14.2 points), with the model collapsing into format imitation and frequent large score deviations; nosvg costs only 0.045 MAE overall, yet Spatial degrades from 0.5167 to 0.6400 and Structural from 0.5840 to 0.6349 — confirming that the rendering dominates perception-based assessment while the code supplies complementary structural cues (path closure, anomalous coordinates, layer ordering) for the SVG-specific dimensions. Beyond the metric drop, nothink also makes the scorer unauditable: correct scores may be incidental, and failures cannot be attributed.
The proposed scorer leads on all four dimensions (MAE 0.5772, 0.13 lower than the best general model), with the largest gains on Spatial and Structural (Spatial 0.5167, the lowest MAE across all models and all dimensions) — indicating that explicit visual grounding plus code cues plus rationale supervision patch precisely what general models lack.

Highlights & Insights¶

Moving the evaluation anchor from "code" to "rendered outcome" is the paper's strongest move. It does not dismiss renderings; it redefines what is being judged: labels look only at the picture, while inputs include the code. This asymmetric design avoids both the blindness of pure code evaluation and the misalignment of pure image metrics, and it is a transferable evaluation paradigm.
"Explicit errors first, Aesthetic as residual" is a cheap and effective disambiguation rule. Multi-aspect scoring naturally suffers from dimension overlap; instead of asking annotators to weigh trade-offs repeatedly, the rule gives them an executable convention that rescues Aesthetic from being a catch-all. The same rule is written into the distillation prompt, so protocol, annotation, and training share one logic — the key to the framework's self-consistency.
Score-conditioned rationale distillation is worth more than score-only supervision. Forcing the teacher to reconstruct evidence for a known gold score makes the "why this score" reasoning explicit; the ablation shows removing it costs 0.256 MAE, and more importantly it preserves auditability — a correct score without evidence cannot be trusted, which is fatal when the evaluator itself is used as a training signal.
The modality-ablation design is worth borrowing: removing the image and the code separately, rather than a single all-vs-text comparison, is what reveals that the code's benefit concentrates on the two vector-specific dimensions. To argue that a modality "provides complementary cues," one must show its gain is not uniform but structurally located where it should matter.
Transferable direction: the same template — "rendering + code dual channel + dimension-wise attribution + evidence distillation" — applies directly to evaluating other structured generative artifacts such as chart generation, LaTeX formula rendering, and CAD/flowchart generation, all of which share the "program representation + human-viewed picture" structure and the same insensitivity of raster metrics to structural defects.

Limitations & Future Work¶

No correlation calibration against existing metrics. The paper argues that MSE/LPIPS/FID/CLIPScore are insensitive to SVG structural defects, yet the experiments only report MAE / Adj. Acc. between model scores and human gold — the most direct counter-evidence, a correlation coefficient between those raster metrics and human scores (or SVGEval scores), is missing. The current argument is analytical (supported by the per-dimension comparisons) rather than controlled-experimental, and this is the most valuable gap to fill.
Gold scores are defined over renderings from a fixed rasterization pipeline. The authors acknowledge that in occasional cases subtle rasterization artifacts may influence model judgments about structural precision. "Rendering as ground truth" buys perceptual alignment while introducing a potential bias source: Structural scores may be affected by the renderer rather than by the vector structure. The authors propose more explicit image–SVG cross-referenced evaluation strategies as the remedy.
Subjectivity is not modeled. Fine-grained judgments on a 1–5 Likert scale, especially for aesthetics, are inherently subjective, and disagreements on borderline cases cannot be fully eliminated. The authors suggest modeling annotations as expected perceptual variance to reduce subjective noise, but this is a direction rather than an implemented mechanism.
Scale and split information are thin. Each part has 500 instances, the prompt pool is 1,000 and the candidate pool about 10k, which is small for a benchmark of this kind; the cached full text also does not specify the train/test split or whether prompt leakage exists (scorer training data and benchmark prompts come from the same source). This directly affects the fairness of the "our scorer vs general models" comparison in Table 2, since general models answer zero-shot while the scorer may have seen same-distribution synthetic data.
Training details for the scorer are missing. The student backbone, teacher model, and the joint rationale + score-token loss are described, but learning rate, epochs, and any anti-overfitting or instruction-mixing measures are not given, raising the reproduction barrier.
Concrete improvements: add a correlation study against LPIPS / FID / CLIPScore and human preferences; release the benchmark and annotation protocol; add a "rank multiple candidate SVGs for the same prompt" task format to reduce the subjectivity of absolute scoring via pairwise comparison.

vs SVG-Bench / StarVector: StarVector proposes task-oriented SVG evaluation (was the generation task completed correctly), whereas SVGEval targets quality diagnosis (is it well drawn, and where does it fail). The former answers "did this prompt produce anything," the latter answers "is the produced vector usable in geometry and layout" — complementary rather than competing.
vs VGBench / SVGenius / VCode / VectorGym: these benchmarks cover multi-task vector pipelines including understanding, editing, and generation, and report pass rates over capability axes; SVGEval instead defines a multi-aspect rubric over rendered outcomes and additionally supports training a stable, evidence-grounded scorer. The difference is a "test paper" versus a "test paper plus a trainable grading model."
vs Q-Bench / Q-Bench+: the Q-Bench line probes low-level visual perception and image-quality understanding, with Q-Bench+ extending to image pairs, so the goal of probing perceptual judgment matches; but their object is natural images, while SVGEval targets SVG renderings and explicitly models spatial layout and structural quality as vector-specific dimensions.
vs G-Eval / Prometheus (rubric-guided evaluators): SVGEval adopts the rubric-guided evaluation philosophy but specializes it: the scorer takes the (prompt, rendering, SVG code) triple rather than text alone, outputs dimension-wise scores with evidence-grounded rationales, and targets SVG-specific spatial and structural defects. In other words, it turns "generic LLM-as-a-judge" into a visual-grounded, dedicated grading pipeline for structured graphics generation.
Insight: the real contribution here may not be "yet another benchmark" but a demonstration of how to design evaluation when a generative modality has a program representation — strip away the surface form of the representation, anchor judgment to the consumption end (the picture a human sees), and keep the representation as verifiable supplementary evidence. Any task on the "program → rendering → human" chain is worth copying this pattern.

Rating¶

Novelty: ⭐⭐⭐⭐ The first framework to anchor SVG quality evaluation on rendered outcomes, with a four-aspect attribution protocol and an explainable scorer; however, rubric-based evaluation and reasoning distillation are transfers of existing paradigms.
Experimental Thoroughness: ⭐⭐⭐ Covers 12 general models plus three modality/supervision ablations, but lacks correlation studies against existing raster metrics and human preferences, and the scale, splits, and training hyper-parameters are not reported.
Writing Quality: ⭐⭐⭐⭐ The evaluation principle (target, four dimensions, hierarchical attribution) is explained clearly and self-consistently; the weaknesses are garbled equations in the extracted version and some missing implementation details.
Value: ⭐⭐⭐⭐ Fills the missing quality-evaluation link for text-to-SVG; the scorer can serve directly as a feedback signal on the generation side, and the "rendering + code dual channel + evidence distillation" design transfers to other structured generation tasks.