RefGlitch Data

This repository provides the RefGlitch data release in the paper https://arxiv.org/pdf/2604.11082.

The data is organized around three pieces:

1. Original videos in video_godot/
2. Released keyframes in keyframes_godot_*
3. Test/reference image pairs in godot_all_1000samples.jsonl

Repository Layout

.
├── video_godot/
│   ├── clipping/
│   ├── floating/
│   ├── lighting/
│   ├── missing_obj/
│   └── texture/
├── keyframes_godot_clipping/
├── keyframes_godot_floating/
├── keyframes_godot_lighting/
├── keyframes_godot_missing_obj/
├── keyframes_godot_texture/
└── godot_all_1000samples.jsonl

All Godot videos are stored under video_godot/, split by bug category:

• video_godot/clipping/
• video_godot/floating/
• video_godot/lighting/
• video_godot/missing_obj/
• video_godot/texture/

Each category contains 100 .mp4 videos.

Released keyframes are stored in the keyframes_godot_* directories:

• keyframes_godot_clipping/
• keyframes_godot_floating/
• keyframes_godot_lighting/
• keyframes_godot_missing_obj/
• keyframes_godot_texture/

Inside each keyframe directory, each released keyframe set lives in a folder named after the corresponding video filename stem, and that folder contains .jpg keyframes for that video.

Examples:

• video_godot/clipping/godot_clipping_029.mp4 -> keyframes_godot_clipping/godot_clipping_029/

Note: we provide the released keyframes that form the test/reference pairs while considering the space constraints of a HuggingFace dataset repository.

The test/reference pairs are stored in:

• godot_all_1000samples.jsonl

This file is a JSONL file with 1000 samples. Each line contains:

• ref_path: path to the reference image
• test_path: path to the test image
• label: 0 for a normal pair, 1 for a glitch pair
• class: glitch type, or normal

All paths in the JSONL are relative to the repository root, so you can load them directly after joining with the local dataset directory.

Minimal Usage Example

import json
from pathlib import Path

root = Path(".")

with open(root / "godot_all_1000samples.jsonl", "r") as f:
    first = json.loads(next(f))

ref_image = root / first["ref_path"]
test_image = root / first["test_path"]

print("Reference:", ref_image)
print("Test:", test_image)
print("Label:", first["label"])
print("Class:", first["class"])

This will give you the image paths for one evaluation pair, which you can then load with your preferred VLMs.