VISTA-Bench

Documentation: English | 简体中文

A multilingual benchmark for evaluating translation from real-world images and released OCR.

Results

Overall denotes global translation quality (T), not a composite of V, T, and K.

Model	Overall	Transport	Dining	Products	Tourism	Healthcare	Safety	Social Media	Publications	Finance	Public Services
Image-input: closed-weight models
GPT-5.6 Luna	55.88	51.00	60.03	53.44	54.38	61.84	60.94	58.92	58.59	58.35	63.55
Claude Sonnet 5	69.43	64.71	71.78	65.71	71.14	74.18	77.94	67.08	71.38	78.13	77.82
Muse Spark 1.3	71.56	64.09	71.98	70.13	71.99	77.43	77.59	73.70	76.86	72.75	73.95
Qwen3.8 Max	72.67	66.00	71.94	72.57	72.89	76.20	81.07	74.67	75.57	76.43	75.40
Gemini 3.8 Flash	74.20	69.10	75.62	74.57	76.22	79.52	79.12	72.64	76.34	78.40	78.17
Doubao Seed 2.1 Pro	75.12	68.43	73.48	75.66	74.24	77.60	81.89	77.48	78.38	80.20	78.87
GPT-6 Astra	77.63	73.78	76.03	76.11	77.02	76.98	83.73	81.63	80.19	79.70	79.17
Image-input: open-weight models
Gemma-4 E4B	19.54	20.55	12.46	17.12	23.60	23.85	27.74	23.53	17.65	15.83	30.80
GLM-4.1V 9B	29.96	30.39	25.36	27.78	31.06	36.42	33.67	31.99	32.86	38.78	35.71
GLM-5.3 Flash	55.33	52.41	55.44	53.89	54.47	63.23	60.37	55.24	60.74	64.68	60.62
DeepSeek Flash	61.04	52.41	62.95	57.80	63.44	69.86	69.91	62.51	65.88	68.92	68.15
Qwen3.8 27B	61.11	56.79	55.63	60.18	63.22	68.80	67.59	63.18	67.32	65.60	68.52
Text-input models
Aya-23 8B	47.43	48.97	41.83	52.95	51.41	54.14	49.67	46.79	46.73	40.91	48.96
Gemma-3 4B IT	55.99	56.22	52.28	60.13	57.92	63.03	58.72	54.22	55.29	50.11	58.52
Qwen3 4B	64.34	65.44	54.00	68.55	66.27	72.38	67.10	61.73	65.56	59.56	67.60
HY-MT2 30B-A3B	72.82	72.75	71.05	78.02	75.65	81.67	75.42	68.12	75.22	64.56	76.15

Overview

VISTA-Bench covers 23 source languages and 10 everyday-use domains. Each image is paired with OCR, verified translations, and two evaluation rule sets: rubric, which retains harder criteria for stronger models, and basic_rubric, whose criteria are intentionally easier for smaller models. A candidate model produces a translation, a text-only judge evaluates it against the target-language rubric, and deterministic scoring produces task scores and aggregates.

🚀 We will continue updating the dataset with more images, languages, and application scenarios.

Quick Start

Python 3.11 or later is required because the scripts use the standard-library tomllib module.

python -m venv .venv
. .venv/bin/activate
python -m pip install -r requirements.txt

The code repository does not contain the dataset. Download data/ from Hugging Face before the first run:

python -m pip install -U huggingface_hub
hf download lvpokoto/VISTA-Bench \
  --repo-type dataset \
  --include "data/**" \
  --local-dir .

Run the benchmark

The easiest option is the interactive launcher. It asks for the API endpoint, models, evaluation track, rubric, target languages, and output directory. API keys entered interactively are hidden and are not written to the repository.

python run_benchmark.py

For a reproducible command-line run, store the API key in an environment variable and start with a small three-task check:

export VISTA_API_KEY='your-secret-value'

python run_benchmark.py \
  --non-interactive \
  --api-url https://api.example.com/v1 \
  --key-env VISTA_API_KEY \
  --candidate-model your-candidate-model \
  --judge-model your-judge-model \
  --endpoint chat/completions \
  --track image \
  --rubric-set rubric \
  --target-lang zh-Hans \
  --limit 3 \
  --output-dir outputs/my-run

After the pipeline completes successfully, remove --limit 3 to evaluate all matching tasks. Repeat --target-lang to select multiple targets. Add --dry-run to inspect the selected tasks without sending API requests.

Unified entry-point options

Option	Meaning
--interactive	Force the setup wizard; it is the default for an argument-free interactive terminal.
--non-interactive	Disable prompts for CI; all required values and the key environment variable must already exist.
--api-url	OpenAI-compatible API base URL; never embed a key in the URL.
--key-env	Name of the environment variable holding the API key, not the key value.
--candidate-model	Model that generates candidate translations.
--judge-model	Text-judge model; it may differ from the candidate model.
--endpoint	chat/completions or responses.
--track	image or text.
--rubric-set	rubric: retains selected hard rules for stronger models; basic_rubric: uses easier individual criteria for smaller or less capable models.
--target-lang	Filter by target language; repeat for values such as zh-Hans and es.
--limit	Run only the first N selected tasks; useful for smoke tests.
--output-dir	Directory for this run's artifacts.
--resume	Continue a compatible output directory and skip completed work.
--dry-run	Validate configuration and print the plan without making API requests.
--candidate-workers	Candidate-stage concurrency.
--judge-workers	Judge-stage concurrency.
--max-attempts	Maximum normal attempts per API stage, from 1 to 3.
--timeout-seconds	Per-request API timeout in seconds.

Candidate and judge can normally share --api-url, --endpoint, and --key-env, while using different values for --candidate-model and --judge-model.

Pipeline and tracks

1. Task selection: select image, source-language, and target-language directions from the current manifest.
2. Candidate: the image track receives the image; the text track receives released OCR only. Both produce self-segmented candidate blocks.
3. Text-only judge: receives source blocks, candidate blocks, target-specific criteria, and tolerances; it never receives the image or reference translation.
4. Scoring: scoring deterministically evaluates the judge checklist.
5. Summary: produces coverage, separate V/T/K direction and domain scores, and independent rankings.

The tracks are separate comparison groups:

• Image track: the candidate sees the image; image understanding, translation quality, and knowledge entailment can be evaluated.
• Text track: the candidate sees OCR only; visual criteria are removed before judging and scoring.

Image and text are separate comparison groups: their scores are not directly comparable and must not be mixed in one leaderboard. Both report the applicable V (visual understanding), T (translation quality), and K (knowledge entailment) dimensions separately, with no cross-dimension weighting or renormalization.

Rubrics

rubric

The primary evaluation collection, recommended for stronger models. It selects and retains hard evaluation rules to differentiate stronger model capabilities.

basic_rubric

The baseline rule set for smaller or less capable models. “Basic” means that each criterion is easier to satisfy, not that the set contains fewer criteria; under the current target-language projections, basic_rubric has more criterion instances than rubric. Its release assets remain under data/simple_rubric/ only to preserve compatibility with the existing data manifest.

The two rubric sets represent different evaluation conditions and must be aggregated and ranked separately; they must never share a leaderboard. A zero-criterion task makes no judge request and is recorded as unscorable; it is neither a numeric zero nor an API failure.

Scoring

• Each expected atom is binary: met = 1 or unmet = 0. Only full satisfaction is met; partial satisfaction must be unmet, and the judge does not accept partial.
• Criterion score = fully met expected-atom count / total expected-atom count. If any forbidden atom matches, that criterion is set to 0.
• V (visual understanding), T (translation quality), and K (knowledge entailment) are scored separately. Within each dimension, criterion importance is weighted as core = 3, supporting = 2, and minor = 1.
• Dimensions have no cross-dimension weights. No composite overall is computed, and missing dimensions are not renormalized.
• The three dimensions are aggregated and ranked independently. One leaderboard.csv stores V/T/K scores and three independent ranks; each rubric set still has its own leaderboard and the two sets must not be mixed.
• Null, failed, and unscorable tasks are excluded; valid numeric zeros remain included.

Outputs and resume

--output-dir contains:

File	Contents
run.json	Key-free run parameters, stage status, and counts; the resume compatibility record.
candidates.jsonl	Self-segmented candidate translations.
candidate_failures.jsonl	Candidate failures for audit and retry.
judgments.jsonl	Per-task checklist results from the text-only judge.
scores.jsonl	Per-task criterion details, V/T/K dimension scores, and statuses; no composite overall.
summary.json	Profile, rubric, coverage, V/T/K aggregation rules, and three independent ranking summaries.
leaderboard.csv	V/T/K scores and three independent ranks in one table; no composite overall rank.
coverage.csv	Success, failure, and unscorable coverage.
scores_by_direction.csv	Separate V/T/K source-to-target aggregates.
scores_by_direction_level1.csv	Separate V/T/K source-to-target × top-level-domain aggregates.
scores_by_level1.csv	Separate V/T/K top-level-domain macros over valid directions.

--resume must point to the original output directory. The entry point reuses the same track, rubric, models, API settings, target languages, and task scope, skips valid candidate/judge records, and regenerates deterministic scores and summaries. Use a new output directory when core parameters change. Do not manually concatenate JSONL from different tracks, rubrics, or models.

Data directories

• data/manifest.jsonl: sample and asset index.
• data/images/, data/ocr/, data/verified_translations/: images, OCR, and verified translations.
• data/simple_rubric/: compatibility directory backing basic_rubric.
• data/rubric/: harder criteria for stronger models.
• script/: candidate, judge, scoring, and summarization scripts plus schemas.

Standalone scripts (advanced)

Most users should run python run_benchmark.py. Copy and edit config.example.toml only when custom intermediate files or pipeline research require the standalone scripts:

cp config.example.toml config.toml
export VISTA_API_KEY='your-secret-value'

python script/generate_image_candidates.py \
  --config config.toml \
  --output outputs/image_candidates.jsonl \
  --failures-output outputs/image_candidate_failures.jsonl \
  --workers 8 --max-attempts 3 --resume

python script/judge.py \
  --config config.toml \
  --candidates outputs/image_candidates.jsonl \
  --rubric-set rubric \
  --profile image \
  --output outputs/image_judgments.jsonl \
  --workers 8 --max-attempts 3 --resume

python script/scoring.py \
  --judge-records outputs/image_judgments.jsonl \
  --rubric-set rubric \
  --profile image \
  --output outputs/image_scores.jsonl

python script/summarize.py \
  --scores model-a=outputs/image_scores.jsonl \
  --profile image \
  --rubric-set rubric \
  --output-dir outputs/image_summary

For the text track, replace the first command with script/generate_text_candidates.py and use --profile text for judge, scoring, and summarization. Never combine profiles or rubric sets in one summary.

Citation

If you find this project useful for your research, please cite:

@misc{lv2026vistabenchbenchmarkingmultilingualimage,
  title={VISTA-Bench: Benchmarking Multilingual Image Translation with Image-Specific Rubrics},
  author={Bo Lv and Mao Zheng and Zheng Li and Fangxu Liu and Mingrui Sun and Tao Chen},
  year={2026},
  eprint={2609.37287},
  archivePrefix={arXiv},
  primaryClass={cs.CV},
  url={https://arxiv.org/abs/2609.37287}
}