The Challenge: 50,000 Images That Need Perfect White Backgrounds
A mid-market fashion retailer operating across the UK and Europe photographs roughly 50,000 product images per month — garments on mannequins, accessories on surfaces, shoes on risers. Every image must ship with a clean white background for the e-commerce site, a lifestyle-style transparent cutout for social media, and a shadow-preserved version for the mobile app. Their outsourced editing studio in Southeast Asia charges £0.35 per image with a 48-hour turnaround. At 50,000 images monthly, that is £17,500 per month and two days of delay before products can go live. Peak season doubles the volume and the studio cannot guarantee turnaround, causing new collection launches to slip.
Cloud-based background removal APIs (remove.bg, Photoroom) charge £0.05–£0.20 per image at volume. Quality on fashion items with fine details — lace, mesh, flyaway hair on model shots — varies, and the retailer has no control over the model or ability to fine-tune for their specific photography style. Sending unreleased product imagery to external APIs also risks competitive data leakage.
AI Solution: Self-Hosted Segmentation Models
State-of-the-art background removal uses segmentation models like RMBG-2.0, U2-Net, IS-Net, or Meta’s Segment Anything Model (SAM). These models generate pixel-precise alpha mattes that separate foreground products from any background. Running the model on a dedicated GPU server enables batch processing of the entire monthly volume in hours rather than days.
The pipeline ingests raw product photographs, runs segmentation to generate alpha masks, composites the cutout onto white/transparent backgrounds, and optionally generates drop shadows using a secondary model. A ComfyUI workflow can orchestrate the entire chain visually, making it accessible to the photography team without coding expertise.
GPU Requirements
Segmentation models range from lightweight (U2-Net at 176 MB) to substantial (SAM-HQ at 2.5 GB). Processing speed depends on input resolution — fashion product images typically arrive at 4000×6000 pixels and are downscaled for inference before the mask is applied at full resolution.
| GPU Model | VRAM | Images per Hour (RMBG-2.0) | Monthly Batch (50K images) |
|---|---|---|---|
| NVIDIA RTX 5090 | 24 GB | ~3,200 | ~16 hours |
| NVIDIA RTX 6000 Pro | 48 GB | ~2,800 | ~18 hours |
| NVIDIA RTX 6000 Pro | 48 GB | ~3,500 | ~14 hours |
| NVIDIA RTX 6000 Pro 96 GB | 80 GB | ~4,200 | ~12 hours |
For a retailer processing 50,000 images monthly, even a single RTX 5090 completes the batch overnight. Running the pipeline during off-peak hours means the GPU is available for other tasks — Stable Diffusion lifestyle image generation, for instance — during the working day.
Recommended Stack
- RMBG-2.0 or IS-Net for background removal, offering the best balance of speed and edge quality on fashion imagery.
- ComfyUI for building visual processing workflows that the photography team can adjust without developer involvement.
- Pillow and OpenCV for post-processing: compositing onto white backgrounds, generating shadows, resizing for web/mobile formats.
- MinIO or local NFS for image storage, with a FastAPI endpoint for triggering batch jobs.
For retailers wanting to go further, pair background removal with an AI image generator to place products into lifestyle scenes — a handbag on a cafe table, trainers on a running track — without a physical photoshoot. A vision model can also auto-tag images with attributes like colour, pattern, and garment type.
Cost Analysis
The outsourced studio costs £17,500 monthly. Cloud APIs at volume pricing run £2,500–£10,000 monthly. Self-hosting on a dedicated GPU server reduces the per-image processing cost to effectively zero after the fixed server fee — a 94% reduction compared to outsourcing and 60-85% compared to APIs. The retailer also gains same-day turnaround instead of 48 hours, accelerating time-to-market for new products.
Additionally, eliminating the two-day editing delay on 50,000 images per month means new products go live sooner, capturing an estimated £22,000 in additional monthly revenue from early visibility during peak browsing periods.
Getting Started
Start with a pilot batch of 500 images across your most challenging product categories — mesh fabrics, jewellery with fine chains, shoes with complex soles. Compare the AI output against your studio’s manual edits. Most retailers find that RMBG-2.0 matches or exceeds human quality on 92-95% of images, with only the most complex cases requiring manual touch-up.
GigaGPU provides UK-based dedicated GPU servers pre-configured for image processing workloads. Add private AI hosting for complete data isolation, or pair with an AI chatbot for automated product photography intake workflows.
GigaGPU offers dedicated GPU servers in UK data centres with full GDPR compliance. Process thousands of product images per hour on private infrastructure.
View Dedicated GPU Plans
