Datasetpublic

8Planetterraforming/geometric_optics_physical_consistency_eval

Geometric Optics – Physical Consistency Evaluation Dataset This dataset contains visual failure cases in geometric optics for multimodal image generation models. Scope The dataset focuses on physical and geometric inconsistencies related to: mirror reflections (law of reflection) refraction and dispersion in prisms light ray direction consistency shadow direction vs light source camera–object–light spatial coherence Motivation Current image… See the full description on the dataset page: https://huggingface.co/datasets/8Planetterraforming/geometric_optics_physical_consistency_eval.

sourceHugging Facecc-by-4.0updated 8mo agoView on Hugging Face

0likes72downloads

Dataset Card

Geometric Optics – Physical Consistency Evaluation Dataset

This dataset contains visual failure cases in geometric optics for multimodal image generation models.

Scope

The dataset focuses on physical and geometric inconsistencies related to:

—mirror reflections (law of reflection)
—refraction and dispersion in prisms
—light ray direction consistency
—shadow direction vs light source
—camera–object–light spatial coherence

Motivation

Current image generation models often produce visually plausible results that violate basic physical laws of optics. These failures persist even with explicit textual constraints.

This dataset is designed for evaluation (not training) of multimodal reasoning and physical consistency.

Failure Categories

—reflectionangleviolation
—refractiondirectionerror
—dispersionorderincorrect
—shadowlightmismatch
—impossiblelightpath

Intended Use

—Model evaluation (GDPval / Evals)
—Multimodal reasoning research
—Physics-aware image generation analysis

License

CC-BY-4.0