---
title: 'VisuGen Metrics: Evaluating Visual Generation'
url: https://www.emergentmind.com/topics/visugen-metrics
type: topic
---

# VisuGen Metrics: Evaluating Visual Generation

Searching arXiv for recent papers explicitly using or defining “VisuGen Metrics” and closely related visualization-generation evaluation metrics.
In the cited literature, **“VisuGen Metrics”** denotes a cluster of evaluation frameworks for generation tasks conditioned on visual inputs or producing visual artifacts. The term is used most explicitly in VisuCraft for the triplet **Visual Grounding (VG), Creativity (C), and Instruction Adherence (IA)**, averaged into a Mean score for long-form creative text generation from images [2508.02890]. Closely related work extends the same evaluative agenda to distribution-aware caption assessment, supervised calibration of caption metrics against human judgments, chart-generation scoring at the specification and rendered-image levels, structural analysis of SVG outputs, and component-wise accuracy measurement for Image→PlantUML conversion [2209.07518], [2410.02381], [2601.15385], [2604.08809], [2509.11667]. Taken together, these frameworks treat evaluation not as a single scalar proxy, but as a decomposition over fidelity, diversity, compliance, structure, and human alignment.

## 1. Terminological scope and evaluative targets

The most direct formalization of **VisuGen Metrics** appears in VisuCraft, where the metric family is defined as **Visual Grounding**, **Creativity**, and **Instruction Adherence**, with a simple arithmetic Mean over the three axes [2508.02890]. In other papers, the same label or a closely related “VisuGen” framing is applied to different output spaces: SVG generation via leave-one-out structural analysis, Image→PlantUML conversion via insertion/deletion/substitution rates, and visual-description evaluation via distribution-aware comparison of caption sets rather than single captions [2604.08809], [2509.11667], [2209.07518].

These frameworks share a common motivation. Traditional evaluation pipelines often reduce performance to pairwise overlap against references or to a single similarity score, which can miss semantically valid diversity, structural modularity, editability, or prompt compliance. Chan et al. explicitly argue that existing metrics are not appropriate for domains such as visual description where reference diversity is signal rather than noise [2209.07518]. A separate audit of image-captioning metrics shows that different metrics measure sharply different properties: n-gram overlap, scene-graph semantics, contextual semantic similarity, and image–text compatibility do not collapse to the same notion of quality [2207.01733].

This suggests that “VisuGen Metrics” is best understood not as a single standardized benchmark, but as a family of task-specific metrics designed to expose failure modes that single-score evaluation suppresses.

## 2. The VisuCraft triad: Visual Grounding, Creativity, and Instruction Adherence

In VisuCraft, **Visual Grounding (VG)** measures how faithfully the generated text $T$ realizes the structured visual attributes $V=\{v_1,\dots,v_M\}$ extracted from an input image. With $\mathrm{match}(v_i,T)=1$ if attribute $v_i$ is realized in $T$ and $0$ otherwise, the score is

$$
VG(T;V) = \frac{1}{M} \sum_{i=1}^{M} \mathrm{match}\bigl(v_i,\,T\bigr)\quad\in[0,1].
$$

The matching procedure uses exact or fuzzy-string matching plus synonym expansion. The paper’s intuition is operational: VG measures whether the output text “mentions” or reflects the fine-grained visual attributes produced by the extractor $\mathcal{E}$, such as “desolate cliff,” “flickering lighthouse beam,” or “stormy gray sea” [2508.02890].

**Creativity (C)** is defined as a weighted combination of lexical diversity and semantic novelty relative to the training corpus. The lexical term is

$$
C_{\mathrm{lex}}(T) = \frac{\bigl|\{\text{distinct }n\text{-grams in }T\}\bigr|}{\bigl|\{\text{all }n\text{-grams in }T\}\bigr|},
$$

and the novelty term is

$$
C_{\mathrm{nov}}(T) = 1 - \max_{R\in\mathcal{D}_{\mathrm{train}}}\cos\!\bigl(f_{\mathrm{text}}(T),\,f_{\mathrm{text}}(R)\bigr).
$$

With default $\lambda=0.5$, the combined score is

$$
C(T) = \tfrac12\,C_{\mathrm{lex}}(T)+\tfrac12\,C_{\mathrm{nov}}(T)\quad\in[0,1].
$$

This construction fixes creativity to two explicit signals: local lexical variation and corpus-level semantic non-redundancy. It does not treat creativity as an unconstrained human-aesthetic judgment; rather, it operationalizes it through diversity and novelty [2508.02890].

**Instruction Adherence (IA)** measures fulfillment of user-specified sub-goals. Given instruction $U$, a parser splits it into $K$ atomic sub-goals $S=\{s_1,\dots,s_K\}$, and the score is

$$
IA(T;U)=\frac{1}{K}\sum_{j=1}^{K}\mathrm{entail}\bigl(s_j,\,T\bigr),
$$

where $\mathrm{entail}(s_j,T)\in\{0,1\}$ is determined by a zero-shot natural-language inference model or, when appropriate, by string or sentiment-analysis checks. The aggregate metric is

$$
\mathrm{Mean}(T)=\tfrac13\bigl[VG(T)+C(T)+IA(T)\bigr].
$$

The evaluation protocol uses **ImageStoryGen-500K**, comprising **500 000 image–instruction pairs**, with **400 k training**, **50 k validation**, and **50 k test** examples across **StoryGen**, **Poetry**, and **AdCopyGen**. Final evaluation samples **5 000 test cases per scenario**, for **15 000 total**, using stratified sampling. Human-in-the-loop calibration includes a validation of the sub-goal parser on **500 instruction texts**, with **92%** end-to-end accuracy, and fuzzy-matching thresholds tuned on a held-out **5K validation subset**, aligning automatic VG with human judgments at **Pearson $r=0.78$** [2508.02890].

The reported results show consistent gains for VisuCraft over baselines. On **StoryGen**, VisuCraft scores **VG 0.825 ± .002**, **C 0.810 ± .002**, **IA 0.830 ± .002**, and **Mean 0.822 ± .002**, compared with **LVLM-Enh Mean 0.811 ± .002** and **LVLM-Base Mean 0.781 ± .003**. On **Poetry**, the corresponding Means are **0.810 ± .003**, **0.794 ± .003**, and **0.772 ± .003**. On **AdCopyGen**, VisuCraft reports **Mean 0.803 ± .003**. All reported differences versus LVLM-Enhanced are significant at **$p<0.001$**, with Cohen’s $d$ for Mean score equal to **0.43** on StoryGen, **0.38** on Poetry, and **0.35** on AdCopyGen [2508.02890].

Ablations further specify what the metric family is sensitive to. Removing $\mathcal{E}$ or $\mathcal{G}$ degrades all three VisuGen Metrics, and richer structured visual granularity monotonically improves **VG**, **C**, **IA**, and **Mean** from **Level 1: Object names** to **Level 3: Full structure**. Reported correlations over all **15 000** samples are **VG vs. IA = 0.82**, **VG vs. C = 0.47**, and **C vs. IA = 0.39**, while failure cases include highly abstract scenes and overly long instructions with more than five sub-goals [2508.02890].

## 3. From single-sample caption scores to distribution-aware visual-description evaluation

A major precursor to later VisuGen-style evaluation is the shift from single-caption comparison to **distribution-aware metrics** for visual description. Chan et al. define an image- or video-conditioned human reference distribution $P_{\mathrm{ref}}(y|x)$ and model distribution $P_{\mathrm{model}}(y|x)$, then compare finite samples

$$
C=\{c_1,\dots,c_N\}\sim P_{\mathrm{model}}(\cdot|x),\qquad
R=\{r_1,\dots,r_M\}\sim P_{\mathrm{ref}}(\cdot|x).
$$

Instead of scoring one candidate against one or more references by max or average overlap, the framework compares the **two distributions** through **Triangle-Rank Metrics (TRM)** and **Kernel-Based Metrics (KBM)** [2209.07518].

For TRM, given a base distance $d(a,b)$, all directed triangles containing at least one sample from $C$ and at least one from $R$ are enumerated. Let $e_{\mathrm{IN}}$ be the edge connecting two samples from the same set. If $I_0(t)$, $I_1(t)$, and $I_2(t)$ indicate whether $e_{\mathrm{IN}}$ is the shortest, middle, or longest edge, then

$$
\mathcal{I}(C,R)=\sum_{t\in T}[I_0(t)+I_1(t)+I_2(t)],
$$

and

$$
Q(C,R)= \left|\frac{\sum_t I_0(t)}{\mathcal{I}(C,R)}-\frac13\right|
+\left|\frac{\sum_t I_1(t)}{\mathcal{I}(C,R)}-\frac13\right|
+\left|\frac{\sum_t I_2(t)}{\mathcal{I}(C,R)}-\frac13\right|.
$$

Under the null hypothesis $P_{\mathrm{model}}=P_{\mathrm{ref}}$, each triangle edge is equally likely to be the in-distribution edge, so each term is near zero; larger $Q$ indicates greater divergence. KBMs include **Fréchet BERT Distance (FBD)** and **MMD-BERT**, using Sentence-BERT embeddings and either Gaussian-distribution comparison or maximum mean discrepancy [2209.07518].

The empirical point is that semantically rich domains, such as image and video description, contain genuine reference diversity. Under single-sample pairwise metrics, models are driven toward a single central caption that overlaps with all references in common n-grams while omitting rarer but valid details. Chan et al. report that under **single-sample evaluation ($N=1$)** no metric can reliably distinguish $P_{\mathrm{model}}$ from $P_{\mathrm{ref}}$ because all **$p>0.05$**, whereas **TRM\_METEOR** and **TRM\_CIDEr** achieve significance much faster as $N$ increases; for example, **TRM\_METEOR reaches $\log p=-3$ at $N=3$ versus $N=7$ for METEOR** [2209.07518].

The framework also changes how quality–diversity trade-offs are interpreted. Standard **METEOR** drops monotonically as temperature increases, but **TRM\_METEOR** peaks at moderate temperature, capturing a regime where diversity improves coverage of reference modes without excessively harming per-sample overlap. Best-practice recommendations in the paper are explicit: report distribution-aware metrics alongside single-sample scores, sample multiple captions with **$N\geq5$**, prefer stochastic decoding over beam search, and use **TRM** on a base metric such as **METEOR** or **CIDEr** for interpretability and analytic $p$-values [2209.07518].

A complementary critique of caption metrics on **MS-COCO 2017 validation** shows why these alternatives matter. In artificial “replacing” and “shuffling” scenarios, **RefCLIPScore** is best on replacing with **Spearman $\rho_s=0.8056$**, while **BERTScore** is best on shuffling with **$\rho_s=0.7096$**. The study concludes that n-gram metrics excel at penalizing random or heavy word substitution but largely fail to detect scrambled word order, whereas CLIP-based metrics capture image–text compatibility and BERTScore captures semantic or fluency degradation [2207.01733]. This directly supports the view that no single conventional caption metric exhausts visually grounded quality.

## 4. Supervised calibration and metric fusion in image captioning

MetaMetrics generalizes the idea of task-specific evaluation by learning a **supervised fusion** over existing metrics to maximize alignment with human judgments. Any base metric $\theta_i$ is treated as a black-box function mapping an input $x$ to a scalar score $\hat{y}_i$, and the meta-metric is defined by a fusion function $\Phi$ over the vector $[\hat{y}_1,\dots,\hat{y}_N]$. In the simplest form,

$$
\hat{y}_{MM}(w)=\sum_{i=1}^{N} w_i\cdot \hat{y}_i,
$$

with weights learned to maximize correlation with human preference scores $z$ [2410.02381].

After preprocessing by clipping, min–max normalization to $[0,1]$, and inversion when higher means worse, the optimization target is

$$
w^*=\arg\max_{w\in W}\rho\!\left(\sum_{i=1}^{N} w_i\hat{y}_i,\; z\right),
$$

where in practice $\rho$ is **Kendall’s $\tau$** on held-out human-rated examples. Under Bayesian Optimization, a Gaussian-process prior is placed on the unknown objective $f(w)=\rho(\hat{y}_{MM}(w),z)$, with a **Matern kernel**. As an alternative, the framework trains an **XGBoost regressor** to predict $z$ from the same feature vector, followed by iterative pruning of the least-important metrics [2410.02381].

For image captioning, **black-Cap** fuses **BLEU-1…4, ROUGE-L, METEOR, SPICE (F/Prec/Rec), CIDEr, TIGEr, and CLIPScore**, including both reference-based and reference-free CLIPScore variants. The vision-task setup uses **Flickr8k-Expert** with **5 882 examples** and a **30%/70%** train/test split, and **THumB 1.0** with **2 000 examples** and the same split. Human labels are on a **1–4** scale for Flickr8k and a **1–5 rubric scale** for THumB. For Bayesian Optimization, the reported hyperparameters are **init\_points=5** and **n\_iter=100**; XGBoost uses **n\_estimators in [100,1 000] (step 100)** with **objective=reg:squaredlogerror** [2410.02381].

The reported improvements are substantial. On **Flickr8k in-dataset**, the best single metric is **SPICE(P) ≈ 0.603**, with **SPICE(F) ≈ 0.591**, **SPICE(R) ≈ 0.579**, and **CLIPScore ≈ 0.514**, whereas **black-Cap (GP)** achieves **$\tau=0.664$** and **black-Cap (XGBoost)** **$\tau=0.670$**. On **THumB 1.0**, the best single metric is **CLIPScore $\tau\approx0.292$**, improved to **0.329** by **black-Cap (GP)**. In **cross-dataset** evaluation, the best individual metric averages about **0.403**, while **black-Cap (XGBoost iterative pruning)** reaches **0.478**, a gain of **0.075**. An ablation shows that the **top-5 metrics retain more than 95% of full-model performance**, and feature-importance maps consistently rely on complementary signals such as **SPICE, CIDEr, CLIPScore**, plus **METEOR** or **TIGEr** as needed [2410.02381].

The broader significance is methodological. MetaMetrics is described as **modality-agnostic and language-agnostic**, with fully parallelizable inference and sparse fusion through automatic pruning. At the same time, the current implementation restricts base metrics to models of at most **10.7B parameters** and avoids “massive” LLM evaluators to ensure **GPU compatibility (≤48 GB)**. Future extensions proposed in the paper include **video captioning/diffusion quality (FID/IS)**, **audio generation**, **larger multilingual/cultural benchmarks**, and more fine-grained aspects of human preference such as **factuality vs style** [2410.02381].

## 5. Specification-level and image-level metrics for visualization generation

For **NL2VIS** and chart generation, VegaChat introduces two complementary metrics: **Spec Score** and **Vision Score**. **Spec Score** compares a generated **Vega-Lite JSON specification** against a reference specification by extracting sets of **mark types**, **encoding channels**, and **data transforms**, then computing precision, recall, and F-scores for each component. Encodings are recall-weighted with $\beta=2$:

$$
F_{\beta,c}=\frac{(1+\beta^2)\cdot(P_c\cdot R_c)}{\beta^2\cdot P_c+R_c}.
$$

With $\beta=2$ for encodings and $\beta=1$ for mark and transform, the final score is

$$
\mathrm{SpecScore}=w_{\mathrm{enc}}\cdot F_{\mathrm{enc}} + w_{\mathrm{mark}}\cdot F_{\mathrm{mark}} + w_{\mathrm{trans}}\cdot F_{\mathrm{trans}},
$$

where the paper chooses the encoding weight highest, for example **0.6**, with **0.2** each for mark and transform, plus a small positive bonus if the spec is valid and non-empty and a heavy penalty if empty [2601.15385].

**Vision Score** is image-based and library-agnostic. A multimodal LLM receives the user prompt, a reference chart image, and a generated chart image, then rates five dimensions on $\{0,1,2\}$: **visualization type**, **data encoding**, **data transformations**, **prompt compliance**, and **aesthetics/layout**. If $s_i$ is the score for dimension $i$ and $w_i$ are manual weights summing to 1, then

$$
\mathrm{VisionScore}=\sum_{i=1}^{5} w_i \cdot \left(\frac{s_i}{2}\right).
$$

The weights place highest emphasis on encoding and lowest on aesthetics. This separates semantic chart correctness from styling while explicitly incorporating prompt compliance, which many chart metrics omit [2601.15385].

The evaluation uses the **NLV Corpus** with **814 single-turn utterances** and **59 sequential sessions**, and an **annotated ChartLLM subset** with **48 real-world chart tasks** and **144 examples total**. On non-sequential NLV, VegaChat reports **Spec Score 83.9%**, **Vision Score 85.1%**, **MPB 61.9%**, **SEVQ 90.6%**, **VER 0.0%**, and **ECR 0.3%**. On ChartLLM, the corresponding values are **Spec 52.6%**, **Vision 56.7%**, **MPB 26.2%**, **SEVQ 80.1%**, **VER 0.0%**, and **ECR 0.8%**. On sequential NLV, **Spec 86.5%**, **Vision 85.7%**, and **ECR 0.0%** are reported. Correlation with human judgments over **171 examples** is **Pearson $r=0.65$** for Spec Score and **$r=0.71$** for Vision Score [2601.15385].

The limitations are equally explicit. **Spec Score** is sensitive to structural mismatches in Vega-Lite JSON and does not handle interactive features or multi-table joins. **Vision Score** depends on an expensive multimodal LLM, tends to assign higher scores in ambiguous cases, and uses manually chosen weights that may not reflect all user preferences. More generally, question-type prompts can admit multiple valid visualizations, so a reference chart may differ from an equally valid generated response [2601.15385].

A useful contrast is provided by **VizGen**, whose evaluation is entirely based on an online **Google Forms survey of $N=19$ participants**. Reported measures are average Likert-scale ratings, such as **Ease of Use 4.4/5.0**, **Confidence Without Support 4.4/5.0**, **Overall UI Design 4.4/5.0**, **Clarity of Demo 4.1/5.0**, **Speed and Responsiveness 3.9/5.0**, and **Navigation Ease 3.3/5.0**, along with feature-adoption percentages and recommendation percentages. The paper explicitly does **not** report **Text-to-SQL translation accuracy**, **graph recommendation precision/recall**, **measured latency**, **insight extraction quality**, or **statistical comparison to baseline tools** [2509.22218]. This suggests why later work emphasizes deterministic, reference-based, and image-based objective metrics for visualization generation.

## 6. Structural and symbolic metrics for SVG and diagram generation

For **text-to-SVG generation**, SVGauge defines a **human-aligned, reference-based metric** by combining **visual fidelity** and **semantic consistency**. Visual fidelity rasterizes the reference SVG $S_O$ and generated SVG $S_G$, extracts SigLIP embeddings, applies PCA/whitening for domain alignment, and computes cosine similarity in the whitened space:

$$
f_{\text{visual}}(S_O,S_G)=\cos(X'_O,X'_G).
$$

Semantic consistency “captures whether the generated SVG truly reflects the prompt’s meaning” by captioning back the generated image with BLIP-2 and comparing prompt and caption in a combined SBERT+TF-IDF space, yielding

$$
f_{\text{semantic}}(p,S_G)=\cos(e(p),e(c)).
$$

The final score is

$$
S_{\mathrm{SVGauge}}(T_O,S_G)=\alpha\,f_{\text{visual}}(S_O,S_G)+(1-\alpha)\,f_{\text{semantic}}(T_O,S_G),
$$

with grid search selecting **$\alpha=0.6$** and **$1-\alpha=0.4$** for maximal Spearman correlation on the held-out set [2509.07127].

The evaluation uses the **SHE benchmark**, containing **333 “gold” SVG images** and their auto-generated prompts, with **2,461 valid SVGs** produced by **eight zero-shot LLM-based SVG generators**. A pool of **40 human annotators** rated how well generated SVGs match prompts on a **1–5 Likert scale**, with each image receiving on average about **8** ratings. At the **system level**, SVGauge reaches **Spearman $\rho=+91.0\%$** and **Kendall’s $\tau=+83.6\%$**, compared with **CLIPScore +80.9\%** Spearman and **+64.2\%** Kendall. At the **instance level**, **SVGauge (ref-based)** reports **Spearman $\rho=+48.4\%$**, versus **CLIPScore +37.8\%** and **SVGauge (ref-free, $\alpha=0$) +36.5\%** [2509.07127].

A second line of work evaluates SVGs structurally through **element-level leave-one-out (LOO) analysis**. Given rendered elements $E=\{e_1,\dots,e_N\}$, the framework renders the full SVG and one ablation per element. With a differentiable reference-based similarity score $S(\cdot,\cdot)$, the **LOO delta** for element $e_i$ is

$$
\Delta_i ::= S(I_{\mathrm{full}}, I_{\mathrm{ref}}) - S(I_i^{-}, I_{\mathrm{ref}}),
$$

and the **pixel-difference mask** is

$$
M_i(x,y) ::= |I_{\mathrm{full}}(x,y)-I_i^{-}(x,y)|.
$$

These signals support **zero-shot artifact detection**, **concept–element attribution**, and four structural metrics: **purity**, **coverage**, **compactness**, and **locality**. The attribution matrix is

$$
A_{ij}=\frac{\sum_{x,y} M_i(x,y)\cdot H_{c_j}(x,y)}{\sum_{x,y} M_i(x,y)+\epsilon},
$$

with $\epsilon=1e^{-6}$, and per-element purity is

$$
\mathrm{purity}(e_i)=\frac{\max_j A_{ij}}{\sum_k A_{ik}+\epsilon}.
$$

At the SVG level, purity averages over active elements; coverage measures the fraction of concepts with at least one assigned element; compactness uses a normalized Herfindahl index over concept attributions; and locality measures how clustered concept-elements are in file z-order [2604.08809].

The validation covers **300 SVGs** across **simple/medium/complex** tiers, **five generation systems**, and roughly **19,000 edits** over **five edit types**. By flagging the **3 most harmful elements** with **$\Delta_i<-0.005$**, the method achieves **F1 ≥ 0.87**, at least **0.17 better than baselines**, and removing them yields **+0.028 SSIM**. On the complex tier, reported purities are **Claude 0.70**, **vtracer 0.68**, **Gemini 0.67**, **GPT-4o 0.66**, **Qwen3 0.64**, and **Source SVG 0.60**. Edit precision tracks purity, with **Claude 0.80**, **vtracer 0.80**, **Gemini 0.79**, **GPT-4 0.76**, **Qwen3 0.74**, and **Source 0.56**. Purity correlates with edit precision at the SVG level with **$r \ge +0.29$, $p<0.001$** for color, delete, move, and scale edits, and **$r=+0.10$, $p=0.02$** for regroup after Bonferroni correction [2604.08809].

For **Image→PlantUML conversion**, the VisuGen framework defines component-wise **Insertion**, **Deletion**, and **Substitution** rates over extracted sets of **nodes**, **arrows/edges**, **messages**, and structural constructs such as **notes**, **boxes**, **groups**, and **participant declarations**. For any class $X$,

$$
I^X = \frac{|X_a\setminus X_{gt}|}{|X_{gt}|},\qquad
D^X = \frac{|X_{gt}\setminus X_a|}{|X_{gt}|},\qquad
S^X = \frac{\left|\{x\in X_a\cap X_{gt}\mid \mathrm{name}_a(x)\neq \mathrm{name}_{gt}(x)\}\right|}{|X_{gt}|}.
$$

Specialized edge metrics include a **Direction-Change Rate** and **Type-Change Rate** for flipped arrow direction and solid-vs-dashed mismatches. The comparison is implemented through **git diff --unified=0**, regex pre-filtering, pairwise Levenshtein distances, and **Jonker–Volgenant** assignment via SciPy’s `linear_sum_assignment` [2509.11667].

The dataset is drawn from **3GPP Release-18 documents**, with about **14,000 images total**, of which **32% (≈4,010)** are sequence diagrams. The evaluation uses a hand-selected subset of **50 diagrams** spanning **1–20** to **51–100** puml lines and totaling about **2,500 ground-truth lines**. On these diagrams, **Claude 3.7 Sonnet** and **GPT-4 Vision** are compared. Claude reports lower average error rates on basic elements; for example, for **Nodes ($\mathcal{N}$)** the insertion/deletion/substitution rates are **13.0 / 15.8 / 12.6**, compared with **19.8 / 18.7 / 34.1** for GPT-4. For **Messages ($\mathcal{M}$)** they are **13.1 / 14.9 / 11.2** versus **19.7 / 17.9 / 39.3**. Both models perform poorly on rare or complex constructs: **GPT-4 deletes all boxes**, while **Claude** reports **42.1 / 52.6 / 21.1** for box insertion/deletion/substitution [2509.11667].

Across these SVG and diagram settings, the common principle is structural accountability. Visual similarity alone cannot identify which elements are harmful, which code fragments support meaningful edits, or which symbolic components were inserted, deleted, or renamed incorrectly. The combined evidence from SVGauge, LOO structural metrics, and PlantUML component-wise rates indicates that human-aligned evaluation for visual generation increasingly depends on decomposing outputs into interpretable substructures rather than treating the artifact as an undifferentiated image [2509.07127], [2604.08809], [2509.11667].

Source: https://www.emergentmind.com/topics/visugen-metrics