LEVI: Benchmarking Visual Identity Fidelity
with Verifiable Low-Entropy Constructs

Xiao Ma¹ Zhonghao Zhan² Hamed Haddadi² Qi Chen^1,*

¹ ZODA² Imperial College London

Hugging Face downloads: coming soon.

A precise failure inside plausible images.

Six identity domains illustrated by correct and incorrect words, geometry, linked rings, piano-key layouts, card counts and drawing stages. — LEVI evaluates the facts that determine visual identity while allowing variation in appearance.
Conceptual illustrations from the paper; actual model outputs appear below.

1,375curated tasks

120base visual identity spaces

6identity domains

10model configurations

Abstract

Modern generative models can produce visually plausible images while violating exact identity constraints. Evaluating such failures requires distinguishing violations of the requested identity from permissible variation in appearance. We introduce Low-Entropy Visual Identity (LEVI), which characterizes visual identities determined by compact facts and explicit rules.

Visual Identity Spaces (VIS) specify admissible identity values and which different realizations express the same identity. Verifiable Visual Grammar (VVG) describes identity-bearing structures through typed primitives and compositional constraints. Together, they support a benchmark that evaluates each generated image against the identity requirements in its text prompt.

The curated benchmark contains 1,375 tasks spanning 120 base visual identity spaces across six domains. Among ten evaluated model configurations, the highest whole-task pass rate is 58.69%, revealing a persistent gap between visually plausible images and faithful realization of their specified identities.

Results

The strongest evaluated configuration passes 58.69% of complete tasks.

Results for ten configurations on the fixed 1,375-task challenge set.
Rank	Model configuration

Loading the paper’s reported results…

Download results JSON

Results from the paper’s retained challenge set, curated after generation. A whole-task PASS requires all evaluated conditions to pass. Domain columns are task-weighted and their task subsets overlap. Generation failures receive zero credit. Labels are automatic evaluation outcomes.

243 / 1,375tasks passed by none of the ten configurations

Success sets also differ across models. Flare passes 155 tasks that Sunburst does not, despite its lower overall score. Aggregate ranking does not imply that one model’s successes contain another’s.

Explore the benchmark

Choose a domain and a task. Compare each model against the same requirement.

What to check

Read the full text prompt

Click any image to enlarge. PASS/FAIL indicate the automatic evaluation outcomes. The schematic illustrates the requested identity and was not supplied to the generators.

Model comparison

Citation

@misc{ma2026levi,
  title         = {{LEVI}: Benchmarking Visual Identity Fidelity with Verifiable Low-Entropy Constructs},
  author        = {Xiao Ma and Zhonghao Zhan and Hamed Haddadi and Qi Chen},
  year          = {2026},
  eprint        = {2610.25716},
  archivePrefix = {arXiv},
  primaryClass  = {cs.CV},
  url           = {https://arxiv.org/abs/2610.25716}
}