LetTokenBurnA benchmark of 1,527 manually screened instances for evaluating single-slide image generation and editing across four tasks with multidimensional metrics.
$ npm i -g lettokenburn-cli$ ltb bench info visslidebench
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 | 8.4 | N/A | Independent | medium | Oct 12, 2026 | arxiv.org |
| Nano Banana 2 | 8.0 | N/A | Independent | medium | Oct 12, 2026 | arxiv.org |
| Qwen-Image-2 | 6.5 | N/A | Independent | medium | Oct 12, 2026 | arxiv.org |
| FLUX.2-dev | 6.4 | N/A | Independent | medium | Oct 12, 2026 | arxiv.org |
| FLUX.2-klein-9B | 6.3 | N/A | Independent | medium | Oct 12, 2026 | arxiv.org |
| HiDream-O1-Image | 4.8 | N/A | Independent | medium | Oct 12, 2026 | arxiv.org |
| BAGEL | 4.4 | N/A | Independent | medium | Oct 12, 2026 | arxiv.org |
| OmniGen2 | 4.1 | N/A | Independent | medium | Oct 12, 2026 | arxiv.org |
Every number links to where it was published. Low-confidence rows are muted and need per-cell verification.
VisSlideBench contains 1,527 instances covering text-to-slide, template-conditioned, multi-image grounded generation, and instruction-based editing. It evaluates content faithfulness, instruction following, aesthetic quality, semantic consistency, and presentation quality.
Scoring uses automatic evaluators validated against blinded human rankings. InfoVis achieves a leading overall score of 7.94 on this benchmark, outperforming other open-source models across eleven metrics.