E-commerce & Retail · Case Study

Brand-consistent product imagery at SKU scale with Flux.1 + a brand LoRA

A multi-brand retailer was waiting weeks for studio photography on every new SKU. We built a Flux.1-based generation pipeline with per-brand style LoRAs, scaled to 12,000 SKUs a day on their existing GPU cluster.

12,000
SKUs/day generated end-to-end
3 weeks → hours
from new SKU to listing-ready image
+18%
conversion lift on regenerated listings (A/B vs old placeholders)
~6s
per 1024×1024 image on H100
Problem

What they were stuck on

The client onboards 800–1,200 new SKUs a week across 14 sub-brands. Studio photography backlogs were running 3 weeks. Stock product shots looked nothing like the brand. Marketplace listings were going live with placeholder images, hurting conversion. They needed studio-quality, on-brand product imagery generated in minutes per SKU - and the look had to stay consistent across batches.

Approach

How we built it

STEP 01

Per-brand LoRA training

Trained 14 Flux.1 dev LoRAs, one per sub-brand. Each LoRA learned the brand's lighting, colour grade, background palette, and shot framing from 200–400 hero shots.

STEP 02

Product-locked composition

ControlNet conditioning on the supplier-provided product cutout, so the actual product geometry is never hallucinated. The LoRA only controls the surrounding scene and styling.

STEP 03

Batch generation API

REST API accepting SKU + brand + scene preset. 4 variations generated per SKU, scored by a fine-tuned aesthetic model, top result auto-promoted.

STEP 04

QC + human-in-the-loop

Confidence-thresholded gating - anything below 0.85 routes to a human reviewer in a custom Streamlit tool. Reviewer feedback flows back into LoRA retraining.

STEP 05

Render farm + CDN

Multi-GPU node with vLLM-style batching for Flux. ~6 seconds per image at 1024×1024 on H100. Generated assets pushed to their existing CDN within 2 minutes of approval.

Stack

What we used

Black Forest LabsFlux.1 devPer-brand LoRA (14×)ControlNetCustom aesthetic scorerH100 render farmStreamlit QC tool
Outcomes

What changed

12,000SKUs/day generated end-to-end
3 weeks → hoursfrom new SKU to listing-ready image
+18%conversion lift on regenerated listings (A/B vs old placeholders)
~6sper 1024×1024 image on H100

We finally have brand-consistent product photography that ships at the pace of merchandising - not at the pace of a studio.

- VP Digital, multi-brand retailer (name withheld)

Have a similar problem? Let's scope it.

A 30-minute call. We'll tell you whether we can help - and if not, who can.

Talk to us