Datasets:

xiaomoguhzz
/

R3_backup

license: apache-2.0
task_categories:
  - image-to-text
  - visual-question-answering
language:
  - en
size_categories:
  - n<1K
tags:
  - benchmark
  - reflection
  - rectification
  - text-to-image
  - visual-reasoning
  - image-editing
pretty_name: R3-Bench

R3-Bench: Reason-Reflect-Rectify Benchmark

Code: github.com/xiaomoguhz/R3-Bench Paper: Benchmarking and Evolving Reason-Reflect-Rectify for Reflective Visual Generation — accepted to ICML 2026.

R³-Bench evaluates reflective visual generation: given a generated image and the original prompt, a model must (i) reason about whether the image matches the prompt, (ii) explain any discrepancy, and (iii) propose a rectification action. The benchmark measures two complementary scores:

S_ref — Reflective Verdict Score (verdict + explanation correctness, judged by an LLM)
S_rect — Rectification Score (normalised VQA-alignment gain after applying the proposed edit)

File	Description
`images.tar.gz`	670 source images (PNG, ~717 MB), organised as `images/{category}/{verdict}/r3b_{idx:06d}.png`

The 670-sample prompt + ground-truth JSONL ships with the code repository at r3bench/data/r3bench.jsonl — labels travel with the eval code so version drift is captured by git.

Schema

Each JSONL record:

{
  "idx": 0,
  "original_prompt": "a black candle and a white holder",
  "bad_image": "images/color/false/r3b_000000.png",
  "answer": false,
  "explanation": "The white object is a candle, not a holder as specified in the prompt.",
  "category": "color",
  "generated_qa": {
    "yn_question_list": ["Is there a candle in the image?", "..."]
  }
}

Field	Type	Description
`idx`	int	Unique sample id (0–669)
`original_prompt`	str	The text-to-image prompt
`bad_image`	str	Image path relative to the data root
`answer`	bool	Ground-truth verdict: does the image match the prompt? (`false` = mismatch)
`explanation`	str	Ground-truth discrepancy description (used as S_ref reference)
`category`	str	Error dimension: color · object · numeracy · spatial · shape · texture · complex · non
`generated_qa.yn_question_list`	list[str]	Yes/no VQA probes used by the S_rect rectification scorer

Category & verdict distribution (670 total)

Category	false (mismatch)	true (match)	Total
color	71	26	97
complex	46	48	94
non	7	42	49
numeracy	73	23	96
object	54	17	71
shape	72	20	92
spatial	76	25	101
texture	49	21	70
Sum	448	222	670

non = "no error" probe samples (image matches prompt) used to calibrate false-positive reflection.

Usage

# 1. Download images
hf download xiaomoguhzz/R3-Bench-data images.tar.gz \
    --repo-type dataset --local-dir /path/to/r3bench-data
cd /path/to/r3bench-data && tar -xzf images.tar.gz
# → /path/to/r3bench-data/images/{category}/{true,false}/r3b_{idx:06d}.png

# 2. Run the 4-step pipeline (clone the code repo first)
git clone https://github.com/xiaomoguhz/R3-Bench && cd R3-Bench
export R3BENCH_DATA_DIR=/path/to/r3bench-data

bash scripts/run_reflection.sh 8 qwen2.5vl    # Step 1 — reflection
bash scripts/run_edit.sh       8 qwen2.5vl qwen_image_2511  # Step 2 — editing
bash scripts/eval_reflection.sh qwen2.5vl       # Step 3 — S_ref
bash scripts/eval_edit.sh      qwen2.5vl qwen_image_2511    # Step 4 — S_rect

See the code repo README for full pipeline details, supported backends, and how to plug in a custom reflection / editor model.

Citation

@inproceedings{r3bench2026,
  title={Benchmarking and Evolving Reason-Reflect-Rectify for Reflective Visual Generation},
  booktitle={Proceedings of the International Conference on Machine Learning (ICML)},
  year={2026}
}

License

Released under the Apache 2.0 license. Source prompts adapted from T2I-CompBench, GenEval++, and GEdit-Bench under their respective licenses; see the code repository for full attribution.