CE-Bench
A 1,200-task benchmark measuring joint satisfaction of task-derived change and preservation requirements in complex image editing.
How to run
$ npm i -g lettokenburn-cli$ ltb bench info ce-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
Scores
12 sourced entries| Harness | Source | Confidence | Link | |||
|---|---|---|---|---|---|---|
| Gemini-3-pro-image-preview | 80.1 | N/A | Independent | medium | Oct 13, 2026 | arxiv.org |
| GPT-Image-1.5 | 77.8 | N/A | Independent | medium | Oct 13, 2026 | arxiv.org |
| DDA-Thinker-32B + Qwen-2511 | 72.9 | dda-thinker-32b | Independent | medium | Oct 13, 2026 | arxiv.org |
| Seedream 4.5 | 71.5 | N/A | Independent | medium | Oct 13, 2026 | arxiv.org |
| EditThinker + Qwen-2511 | 68.5 | editthinker | Independent | medium | Oct 13, 2026 | arxiv.org |
| Gemini-2.5-Flash-Image | 68.1 | N/A | Independent | medium | Oct 13, 2026 | arxiv.org |
| ThinkRL-Edit | 67.8 | dars config. | Independent | medium | Oct 13, 2026 | arxiv.org |
| PromptEnhancerV2 + Qwen-2511 | 66.8 | promptenhancerv2 | Independent | medium | Oct 13, 2026 | arxiv.org |
| FLUX.2-klein-9B | 60.8 | N/A | Independent | medium | Oct 13, 2026 | arxiv.org |
| LongCat-Image-Edit | 60.5 | N/A | Independent | medium | Oct 13, 2026 | arxiv.org |
| RePlan + Qwen-2511 | 59.3 | replan | Independent | medium | Oct 13, 2026 | arxiv.org |
| FLUX.1-Kontext-dev | 43.1 | N/A | Independent | medium | Oct 13, 2026 | arxiv.org |
Every number links to where it was published. Low-confidence rows are muted and need per-cell verification.
Reproduced by the community
All runsAbout
CE-Bench comprises 1,200 tasks designed to evaluate complete compliance with source-grounded changes and preservation constraints in image editing. It assesses whether models faithfully interpret instructions and execute edits without altering protected content.
Scoring relies on a Complete Compliance Rate (CCR), requiring all task-derived checks to pass. The benchmark demonstrates that coordinated understanding and generation improve compliance by over 21 percentage points compared to direct editors.