builditwithgk/adresyze-ad-layouts

AdResyze — Indian Ad Layout Annotations

289 annotated Indian brand advertisements with element bounding boxes, type labels, dominant colors, aspect ratios, and platform guesses.

This is the training dataset used to fine-tune builditwithgk/adresyze-lora — a LoRA adapter for Qwen2.5-VL-7B that understands advertisement layout anatomy.

Dataset Description

Note: This dataset contains annotation JSON files only. The original advertisement images are not redistributed as they belong to the respective brands. Images were collected from Meta Ad Library (India) for research purposes.

Collection Methodology

• Source: Meta Ad Library (facebook.com/ads/library)
• Country: India
• Media type: Image
• Search terms: sale, offer, brand, new launch, discount, shop now, free delivery, cashback
• Collection date: Early 2026
• Total collected: 302 images
• Successfully annotated: 289 images (95.7% success rate)

Annotation Format

Each JSON file corresponds to one advertisement image and contains:

{
  "elements": [
    {
      "type": "logo|headline|cta|product|background|other",
      "bbox": [x1, y1, x2, y2],
      "priority": 1,
      "must_preserve": true
    }
  ],
  "dominant_colors": ["#hex1", "#hex2"],
  "aspect_ratio": "1:1|9:16|4:5|1.91:1|other",
  "platform_guess": "instagram|facebook|linkedin|google_display|other",
  "image_file": "original_filename.jpg"
}

bbox format: [x1, y1, x2, y2] — top-left to bottom-right pixel coordinates.

Element Types

Type	Description
logo	Brand logo or icon
headline	Main text headline or tagline
cta	Call-to-action button or text
product	Product image or visual
background	Background element (full image bounds)
other	Other visual elements

Dataset Statistics

Metric	Value
Total annotations	289
Annotation method	Auto-annotated using Qwen2.5-VL-7B-Instruct
Avg elements per image	3–5
Aspect ratios	1:1 (majority), 1.91:1, 4:5, other
Platform distribution	Instagram, Facebook, LinkedIn, other

Usage

This dataset was used to train the AdResyze LoRA adapter:

import json, os

annotations_dir = "path/to/adresyze-ad-layouts"

for fname in os.listdir(annotations_dir):
    if fname.endswith(".json"):
        with open(f"{annotations_dir}/{fname}") as f:
            annotation = json.load(f)
        
        elements = annotation["elements"]
        aspect_ratio = annotation["aspect_ratio"]
        platform = annotation["platform_guess"]
        
        print(f"File: {fname}")
        print(f"Elements: {len(elements)}")
        print(f"Aspect ratio: {aspect_ratio}")
        print(f"Platform: {platform}")

Related Resources

• Model: builditwithgk/adresyze-lora
• Project: AdResyze — AI-powered ad format resizing

License

CC BY 4.0 — Attribution required. Commercial use permitted.

When using this dataset, please cite:

@misc{adresyze-dataset-2026,
  author = {builditwithgk},
  title  = {AdResyze Indian Ad Layout Annotations},
  year   = {2026},
  url    = {https://huggingface.co/datasets/builditwithgk/adresyze-ad-layouts}
}

Limitations

• Annotation quality depends on base Qwen2.5-VL-7B model — approximately 95% accuracy
• Bounding boxes are approximate, not pixel-perfect
• Images not included — annotations only
• Focused on Indian market brand advertisements