RayyanAhmed9477/pexels-advertisement-photography

Pexels Advertisement & Product Photography Dataset

A curated dataset of 12,917 high-quality advertisement and product photography images sourced from Pexels, enriched with AI-generated captions for text-to-image model fine-tuning.

Dataset Overview

Attribute	Value
Total Images	12,917
Source	Pexels API
Categories	40 search queries
Avg Resolution	842 x 886 px
Min Resolution	512 px (shortest side)
Total Size	~2.4 GB (parquet with embedded images)
Captions	Dual (long descriptive + short tag-style)
Caption Model	Qwen3-VL-8B

Pipeline

This dataset was built through a multi-stage pipeline:

1. Scraping (Pexels API)

• 40 curated search queries spanning product photography, commercial advertising, luxury brands, food, fashion, automotive, tech, and more
• Up to 400 images per query via paginated API calls
• 12,927 unique images downloaded at medium resolution (940px wide)

2. Filtering

• Resolution filter: Minimum 512px on shortest side (removes low-quality thumbnails)
• Perceptual deduplication: pHash-based near-duplicate removal (hamming distance threshold)
• Aesthetic scoring: CLIP-based aesthetic predictor scores each image
• Result: 12,917 images retained (99.9% pass rate — Pexels quality is inherently high)

3. Caption Enrichment (VLM)

Each image was captioned by Qwen3-VL-8B with two complementary styles:

• caption_long: Detailed advertising description (~50-100 words) describing composition, lighting, mood, product placement, and marketing context
• caption_short: Concise tag-style caption (~10-20 words) with comma-separated descriptors for style, category, and key visual elements

This dual-caption approach enables stochastic caption sampling during training (e.g., 70% long / 30% short) for better prompt generalization.

Search Queries (40 Categories)

Click to expand full query list

Product Photography:

• product photography studio white background
• cosmetics product shot professional
• perfume bottle luxury commercial
• jewelry product photography
• luxury watch product shot
• skincare product minimal
• running shoes product
• headphones technology product
• handbag luxury product
• sunglasses product photography
• wine bottle product shot
• smartphone product photography
• beauty product flat lay
• e-commerce product listing

Food & Beverage:

• food photography commercial
• coffee cup latte art
• chocolate dessert food photography
• beverage drink commercial
• food advertisement menu

Automotive & Fashion:

• car automotive photography
• automobile car advertisement
• fashion accessories product
• fashion advertisement editorial
• furniture modern interior design

Advertising & Commercial:

• advertisement commercial campaign
• brand advertisement poster
• luxury brand advertisement
• retail store advertisement display
• billboard advertisement commercial
• magazine advertisement layout
• digital advertisement banner
• commercial branding photography
• marketing campaign visual
• sale promotion advertisement
• shopping advertisement retail
• real estate advertisement property
• tech gadget advertisement
• fitness sports advertisement
• beauty cosmetics advertisement campaign
• holiday seasonal advertisement

Dataset Statistics

Metric	Value
Aesthetic Score (mean)	3.12
Aesthetic Score (median)	3.14
Aesthetic Score (range)	2.53 - 3.55
Width (mean)	842 px
Height (mean)	886 px
Avg Image Size	179 KB

Usage

from datasets import load_dataset

# Load the dataset
ds = load_dataset("RayyanAhmed9477/pexels-advertisement-photography", split="train")

# Access an example
example = ds[0]
image = example["image"]          # PIL Image
long_caption = example["caption_long"]   # Detailed advertising description
short_caption = example["caption_short"] # Tag-style caption
score = example["aesthetic_score"]       # Aesthetic quality score

# For LoRA fine-tuning with stochastic captioning
import random
caption = long_caption if random.random() < 0.7 else short_caption

Use with Diffusers

This dataset is designed for fine-tuning text-to-image diffusion models (e.g., Stable Diffusion, FLUX, Qwen-Image) via LoRA:

from datasets import load_dataset

ds = load_dataset("RayyanAhmed9477/pexels-advertisement-photography", split="train")

# Example: prepare for diffusers training
for example in ds:
    image = example["image"]
    caption = example["caption_long"]  # or caption_short
    # ... your training pipeline

Schema

Column	Type	Description
image	bytes	JPEG image data
text	string	Original Pexels search query used to find this image
dimensions	string	Image dimensions as "WxH"
aesthetic_score	float32	CLIP aesthetic predictor score
caption_long	string	Detailed advertising-style description from Qwen3-VL-8B
caption_short	string	Concise tag-style caption from Qwen3-VL-8B

License & Attribution

• Images are sourced from Pexels under the Pexels License
• Pexels License permits free use for commercial and non-commercial purposes
• No attribution required, but appreciated
• AI-generated captions are provided as-is

Built With

• Pexels API for image sourcing
• Qwen3-VL-8B for caption generation
• CLIP for aesthetic scoring
• pHash for deduplication

Citation

@dataset{pexels_ad_photography_2026,
  title={Pexels Advertisement & Product Photography Dataset},
  author={Rayyan Ahmed},
  year={2026},
  url={https://huggingface.co/datasets/RayyanAhmed9477/pexels-advertisement-photography},
  note={12,917 images with dual AI captions for text-to-image fine-tuning}
}