Datasets:

rosebudtse
/

CompGen-GRPO-eval-samples

license: mit
task_categories:
  - text-to-image
tags:
  - t2i-compbench
  - compositional-generation
  - grpo
  - ablation
size_categories:
  - 100K<n<1M

CompGen-GRPO Eval Samples

Pre-generated images from a series of CompGen-GRPO checkpoints (fine-tuned Janus-Pro-1B with GRPO), evaluated on T2I-CompBench (color / shape / texture / spatial / non_spatial / complex).

Every image is a 384×384 PNG rendered with the same sampling config so results are directly comparable:

Prompt set: T2I-CompBench eval split (300 prompts × 6 categories)
Samples per prompt: 10
Resolution: 384×384
CFG scale: 5.0
Seed: 42
Reasoning prompt: Bi-CoT template (see companion code repo)

Total: ~162,000 images across 9 models × 6 categories × 3000 imgs.

Repo layout

CompGen-GRPO-eval-samples/
├── baseline/            # Janus-Pro-1B (no RL, deepseek-ai release)
│   ├── color/samples.tar
│   ├── shape/samples.tar
│   └── ...
├── full_400/
├── full_800/
├── wo_gdino_400/        # ablation: no GroundingDINO
├── wo_gdino_600/
├── wo_gdino_800/
├── wo_attr_400/         # ablation: no VLM Attr
├── wo_attr_600/
├── wo_attr_800/
└── <MODEL>/<CATEGORY>/samples.tar

Each samples.tar contains a samples/<PROMPT_ID>_<SAMPLE_ID>.png directory (uncompressed tar — PNG is already compressed).

Categories:

Category	Prompts	Total imgs	Eval metric
`color`	300	3000	BLIP-VQA
`shape`	300	3000	BLIP-VQA
`texture`	300	3000	BLIP-VQA
`spatial`	300	3000	UniDet-2D
`non_spatial`	300	3000	CLIPScore
`complex`	300	3000	3-in-1 (BLIP+UniDet+CLIP)

Naming convention: <PROMPT_ID>_<SAMPLE_ID>.png — PROMPT_ID ∈ [0, 300), SAMPLE_ID ∈ [0, 10).

Note on `full_400`

For historical reasons full_400/{color,shape,texture}/samples/ contains 6000 images (20 samples / prompt) instead of the standard 3000. This does not change the mean score meaningfully (n=3000 is already low-variance), but be aware if you rely on per-sample paired comparisons.

How to download

pip install -U huggingface_hub

# Full dataset (all zips)
hf download <HF_USER>/CompGen-GRPO-eval-samples \
    --repo-type dataset \
    --local-dir ./eval_samples

# Just one model
hf download <HF_USER>/CompGen-GRPO-eval-samples \
    --repo-type dataset \
    --include "full_800/*" \
    --local-dir ./eval_samples

# Unzip all
find ./eval_samples -name "samples.tar" | while read t; do
    (cd "$(dirname "$t")" && tar -xf "$(basename "$t")" && rm "$(basename "$t")")
done

How to reproduce eval

git clone https://github.com/xiezifan/CompGen-GRPO
cd CompGen-GRPO
# After unzip, samples land at <MODEL>/<CATEGORY>/samples/*.png — just move
# that tree under eval_results/ and run:
bash src/t2i-r1/src/run_eval.sh --model full_800,wo_gdino_600 --task all --gpu 0

See run_eval.sh for the full T2I-CompBench eval pipeline (BLIP-VQA + UniDet + CLIPScore + 3-in-1).

License

Images are released under CC-BY-4.0. See paper for detailed attribution.

Citation

@article{compgen-grpo-2026,
  title  = {Compositional Text-to-Image Generation via Multi-Reward GRPO},
  author = {Xie, Zifan and others},
  year   = {2026},
  note   = {In preparation}
}