MMMG: A Multi-discipline Benchmark for Professional Text-to-Image Generation

MMMG is a Massive Multi-discipline Multimodal Generation and Reasoning Benchmark for evaluating text-to-image models in professional domains. It covers eight disciplines and was constructed through a five-stage pipeline including domain definition, prompt generation, solution reasoning, checklist construction, and human verification. The dataset is 5.5 MB in size, authored by Li Jiang, and was last updated on 2026-05-19.

Use Cases

• Benchmarking text-to-image models on knowledge-intensive tasks based on the described eight professional disciplines.
• Evaluating the joint capability of logical reasoning and graphical generation based on the benchmark's stated purpose.
• Developing checklist-based evaluation protocols for fine-grained assessment of image generation models.
• Studying model performance on rule-constrained and reasoning-intensive professional scenarios.

Strengths

• Covers eight distinct professional disciplines for domain-specific evaluation.
• Constructed via a detailed five-stage pipeline including human verification.
• Includes a proposed checklist-based evaluation protocol for interpretable assessment.
• File size is 5.5 MB, indicating a structured, likely non-image, dataset.

Limitations

• Row count is unknown, which may limit suitability assessment for large-scale training.
• Column-level documentation is absent; field semantics must be inferred after download.
• Data may reflect bias inherent to the specific construction and verification pipeline described.

Provenance

Source

figshare

Collection Method

Constructed through a unified five-stage pipeline including domain definition, prompt generation, solution reasoning, checklist construction, and human verification.

Freshness

Last updated 2026-05-19 07:21:25; freshness should be verified.