Primer Intelligent Image Processing Pipeline

AI-Powered Automation with Qwen VLM for E-Commerce Assets

An intelligent, high-volume image processing pipeline engineered to automate the resizing, padding, and optimization of product images for e-commerce platforms. Leveraging Qwen Vision-Language Model (VLM) for edge detection and sophisticated algorithms to ensure visual consistency across 25,000+ product assets.

25,000+
Assets Processed
95%
Time Reduction
100%
Visual Consistency

Project Overview

This system is an AI-powered image processing pipeline designed to eliminate manual editing bottlenecks in e-commerce product photography. By combining Qwen VLM for intelligent edge detection with context-aware padding algorithms, it ensures marketplace-ready images at scale while maintaining perfect visual consistency.

Key Achievements

  • ✅ Processed 25,000+ product images across 80+ brands
  • ✅ Reduced manual editing time by 95% (hours → minutes)
  • ✅ Achieved 100% visual consistency across catalogs
  • ✅ Zero manual intervention for standard photography
  • ✅ 98%+ edge detection accuracy with Qwen VLM

Primary Use Cases

  • 🖼️ Automated canvas resizing with intelligent padding
  • 🎯 AI-powered edge detection for cropped products
  • 🔄 Multi-format conversion (RAW, AVIF, CMYK→RGB)
  • ✂️ Background removal with 4 provider options
  • 📐 80+ brand-specific canvas configurations
  • 📦 Batch processing for high-volume workflows

Key Features

Smart Edge Detection

Qwen VLM analyzes product boundaries to identify cropped edges (top/bottom/left/right). Applies context-aware padding rules: zero padding on cropped sides, glued placement for natural appearance.

Adaptive Padding Engine

Intelligent padding algorithms adjust based on edge analysis. Single edge cropped → glue to edge. Multiple edges → adaptive rules. All intact → centered placement with full padding.

Multi-Format Support

Handles RAW files (CR2, NEF, ARW), AVIF, CMYK with ICC profiles. Automatic conversion to web-ready RGB/PNG. Preserves color accuracy through professional color management.

Background Removal

Four provider options: Rembg (local, fast), BIREFNET (GPU, highest quality), PhotoRoom (cloud, professional), Recraft (scalable). Original RGB values preserved after removal.

80+ Canvas Configs

Brand-specific canvas sizes and padding for Lazada/Shopee, DOTCOM, and Zalora platforms. Pre-configured for Aetrex, Allbirds, Birkenstock, Dr. Martens, North Face, and 75+ more brands.

Batch Processing

Concurrent processing with ThreadPoolExecutor. ZIP archive support for bulk uploads. Real-time progress tracking. Processes 50-200 images/minute depending on provider and hardware.

Technology Stack

AI/ML Models

Qwen VLM Rembg (U2-Net) BIREFNET PhotoRoom API Recraft API Transformers

Image Processing

Pillow (PIL) OpenCV rawpy NumPy ImageCms pillow-avif

Backend & UI

Python Gradio PyTorch ThreadPoolExecutor dotenv

Results & Business Impact

Quantified Results

  • 25,000+ product images processed across multiple brands
  • 95% reduction in manual editing time (hours → minutes)
  • 100% visual consistency across e-commerce catalogs
  • 98%+ edge detection accuracy with Qwen VLM
  • 50-200 images/minute batch processing throughput
  • Zero manual intervention for standard product photography

Technical Highlights

  • 80+ brand configurations with platform-specific sizing
  • Multi-format support (RAW, AVIF, CMYK with ICC profiles)
  • 4 background removal providers for quality/speed tradeoffs
  • Context-aware padding adapts to cropped product edges
  • Original color preservation maintains product accuracy
  • Concurrent processing with ThreadPoolExecutor

Interested in Intelligent Image Processing?

Let's discuss how AI-powered automation can transform your e-commerce workflows

← Back to Portfolio Contact Me →