LetTokenBurnUltraText Bench: A Comprehensive Bilingual Benchmark for Evaluating Visual Text Rendering in Image Generation
A bilingual benchmark with 432 prompts across 24 scene categories and 3 difficulty levels, evaluating visual text rendering fidelity, clarity, spatial quality, and scene quality in image generation.
$ npm i -g lettokenburn-cli$ ltb bench info ultratext-bench
External harness. No Harbor adapter has been generated for this benchmark yet. `ltb bench info` prints the known links.
Runs happen on your machine with your own subscription; only the score summary is pushed. CLI docs
| Harness | Source | Confidence | Link | |||
|---|---|---|---|---|---|---|
| GPT-Image-2 | 99.3 | N/A | Independent | medium | Oct 7, 2026 | arxiv.org |
| Nano Banana 2 | 97.0 | N/A | Independent | medium | Oct 7, 2026 | arxiv.org |
| Seedream 5.0 Pro | 91.5 | N/A | Independent | medium | Oct 7, 2026 | arxiv.org |
| Qwen-Image-2.0-Pro | 90.2 | N/A | Independent | medium | Oct 7, 2026 | arxiv.org |
| Qwen-Image-3.0 | 88.3 | N/A | Independent | medium | Oct 7, 2026 | arxiv.org |
| WAN-2.7-Image-Pro | 85.7 | N/A | Independent | medium | Oct 7, 2026 | arxiv.org |
| Boogu-Image-0.1-Base | 77.9 | N/A | Independent | medium | Oct 7, 2026 | arxiv.org |
| Qwen-Image-2512 | 68.0 | N/A | Independent | medium | Oct 7, 2026 | arxiv.org |
| Z-Image-Base | 61.9 | N/A | Independent | medium | Oct 7, 2026 | arxiv.org |
| GPT-Image-1.5 | 58.0 | N/A | Independent | medium | Oct 7, 2026 | arxiv.org |
| HiDream-O1-Image | 45.6 | N/A | Independent | medium | Oct 7, 2026 | arxiv.org |
Every number links to where it was published. Low-confidence rows are muted and need per-cell verification.
Contains 432 human-reviewed prompts spanning 24 real-world scene categories, split equally between English and Chinese. Each prompt specifies exact strings for 4-12 text regions with structured references for content, placement, and attributes.
Uses the Q-Judger VLM to assess images against references, reporting text fidelity, clarity, spatial quality, and scene quality. Evaluation of 24 models reveals trade-offs, such as Z-Image-Turbo gaining clarity but losing fidelity compared to its base version.