RIG-Bench

RIG-Bench is a benchmark for Reasoning-driven Image Generation. Each sample gives a model visual context, a natural-language instruction, and optional demonstration pairs; the model must infer the underlying rule, relation, concept, or process and generate the answer as a single image.

• Paper: Thinking_in_Pictures__A_Systematic_Benchmark_for_Reasoning_driven_Image_Generation.pdf
• License: CC BY-NC 4.0

Dataset Summary

• 2,000 test samples
• 4 task families and 11 fine-grained subtasks
• English instructions
• Image inputs and image targets
• Unified JSONL schema with per-sample image folders
• Released under CC BY-NC 4.0

Task Format

Each sample contains:

• input.images: ordered visual context images
• input.text: task instruction
• input.examples: optional input-output demonstration pairs
• output.target_image: ground-truth answer image
• main_family: high-level task family
• subtask: fine-grained reasoning subtask

The model should infer the answer from the visual context and generate one standalone answer image. Multiple-choice options from the original source tasks are not exposed to the model.

Task Families

Family	Subtask	# Samples
Concept-based	interaction_concept	216
Concept-based	abstract_concept	152
Transformation-based	geometric_analogy	240
Transformation-based	attribute_analogy	240
Transformation-based	rule_transfer	109
Pattern-Structure	matrix_reasoning	326
Pattern-Structure	visual_spatial_reasoning	131
Scenario-based	scientific_process	96
Scenario-based	scene_temporal_inference	180
Scenario-based	cross_domain_analogy	180
Scenario-based	style_preference	130
Total		2,000

RIG-Bench also varies the constraint of the answer space. Open-form subtasks allow multiple visually valid realizations of an inferred concept, style, or scene continuation; closed-form subtasks require a rule-constrained target determined by a geometric, structural, scientific, or analogy relation.

Files

RIG-Bench/
|-- README.md
|-- samples.jsonl
`-- images/
    `-- <sample_id>/
        |-- input_<order>.<ext>
        |-- demo_<j>_input_<k>.<ext>    # present for samples with demonstrations
        |-- demo_<j>_output_<k>.<ext>   # present for samples with demonstrations
        `-- target.<ext>

Schema

Each line of samples.jsonl is a JSON object:

{
  "sample_id": "0001",
  "input": {
    "images": [
      {"path": "images/0001/input_0.jpg", "order": 0}
    ],
    "text": "Generate a new scene that preserves the shared human-object interaction concept.",
    "examples": []
  },
  "output": {
    "target_image": "images/0001/target.jpg"
  },
  "main_family": "Concept-based",
  "subtask": "interaction_concept"
}

Loading Locally

import json
from pathlib import Path

from PIL import Image

root = Path("RIG-Bench")
records = [
    json.loads(line)
    for line in (root / "samples.jsonl").open(encoding="utf-8")
]

sample = records[0]
context_images = [
    Image.open(root / image["path"])
    for image in sample["input"]["images"]
]
target_image = Image.open(root / sample["output"]["target_image"])
instruction = sample["input"]["text"]

For samples with demonstrations, input.examples contains paired demonstration inputs and outputs that should be treated as part of the prompt context.

Intended Use

RIG-Bench is intended for research on multimodal reasoning, image generation, and diagnostic evaluation of reasoning-to-generation failures. It is not intended as a general-purpose training corpus or as evidence of deployment readiness for safety-critical, medical, legal, financial, identity-recognition, surveillance, or demographic-inference applications.

Citation

If you use RIG-Bench, please cite the accompanying paper:

@article{rigbench2026,
  title = {Thinking in Pictures: A Systematic Benchmark for Reasoning-driven Image Generation},
  year = {2026}
}