Arklin123/CV-Arena-Sample

CV-Arena — Anonymous Data Release (NeurIPS 2026 Submission)

Paper: CV-Arena: An Open Benchmark for Instructional Computer Vision Problem Solving with Human-AI Collaborative Preferences

---

About this release

This directory is a sample release accompanying our anonymous NeurIPS 2026 submission. It is not the full CV-Arena benchmark. We provide a small, curated slice so reviewers and readers can inspect the data format, prompt style, and task coverage without downloading the full corpus.

• 10 (image, English-prompt) pairs per subtask.
• 16 subtasks spanning low-level restoration, weather degradation removal, scene editing, face/fashion/virtual-try-on, semantic reconstruction, and text/typography manipulation.
• Images and prompts are taken directly from the full CV-Arena dataset; only the quantity is reduced for this anonymous preview.
• The full dataset, evaluation protocol, and human-AI preference annotations will be released upon acceptance / de-anonymization.

Directory layout

opensource_data/
├── 01_upscaling/
├── 02_weather_fog/
├── 03_weather_rain/
├── 04_weather_snow/
├── 05_blur/
├── 06_old_photo/
├── 07_lowlight/
├── 08_overexposure/
├── 09_scene_composition_and_object_insertion/
├── 10_Face_Attribute_Manipulation/
├── 11_fashion_based_edit/
├── 12_outpainting/
├── 13_semantic_aware_content_reconstruction/
├── 14_Text-based_Geometric_Warping/
├── 15_Typography_UI_Restoration/
└── 16_virtual_try_on/

Each subtask folder contains:

• 1.<ext> … 10.<ext> — 10 source images (.jpg / .png).
• prompts.json — list of records mapping each image to its English instruction prompt.

prompts.json format

[
  {
    "id": 1,
    "image": "1.jpg",
    "prompt": "remove fog in the image and enhancing its quality."
  },
  ...
]

Subtask overview

#	Subtask	Task type
01	upscaling	Super-resolution / detail enhancement
02	weather_fog	Fog removal
03	weather_rain	Rain removal
04	weather_snow	Snow removal
05	blur	Deblurring
06	old_photo	Old-photo colorization / restoration
07	lowlight	Low-light enhancement
08	overexposure	Highlight / over-exposure correction
09	scene_composition_and_object_insertion	Physically-plausible object insertion
10	Face_Attribute_Manipulation	Fine-grained facial attribute editing
11	fashion_based_edit	Clothing / accessory edits
12	outpainting	Context-aware scene outpainting
13	semantic_aware_content_reconstruction	Pose / state / structure transitions
14	Text-based_Geometric_Warping	Geometric / perspective edits driven by text
15	Typography_UI_Restoration	Text / sign / typography repair
16	virtual_try_on	Garment swap / virtual try-on

How to use

import json
from pathlib import Path
from PIL import Image

root = Path("opensource_data")
for subtask_dir in sorted(root.iterdir()):
    if not subtask_dir.is_dir():
        continue
    records = json.load(open(subtask_dir / "prompts.json"))
    for rec in records:
        img = Image.open(subtask_dir / rec["image"])
        prompt = rec["prompt"]
        # ... feed (img, prompt) to your model

Anonymity & licensing

• This release contains no author identifiers, no internal paths, and no preference / annotation data.
• Source images are aggregated from publicly available web data for research benchmarking. Each image is the property of its original creator; we redistribute only for the purpose of academic evaluation under fair-use review.
• Please do not redistribute this sample outside the review context. The official, fully-licensed release will accompany the camera-ready paper.

Contact

To preserve double-blind review, please direct questions through the OpenReview submission page rather than by email.