Fashion Brand Lora

LoRA Training

Fashion Brand LoRA

A custom LoRA model trained on Uzurii’s luxury footwear aesthetic — enabling consistent, on-brand AI image generation across every product style, colourway, and campaign format, without manual iteration.

Client

Uzurii Luxury Footwear

Tools Used

Flux, NanaBanana, ComfyUI

Year

2025

← Back to Portfolio
Uzurii Fashion Brand LoRA — Sample Output

Project Overview

The Challenge

Producing consistent AI imagery for Uzurii with standard prompting had a fundamental limitation: every image required significant manual intervention to accurately reproduce the specific characteristics of their products — the precise material finishes, embellishment details, heel shapes, and the distinctive design language that makes Uzurii recognisable. Even with careful prompt engineering, outputs would drift in style between sessions and lose the brand specificity that luxury clients demand. What was needed wasn’t better prompts — it was a model that had Uzurii’s aesthetic baked in at the weights level.

The Training Process

We assembled a curated dataset of 500+ high-quality reference images covering the full Uzurii product range — carefully selected to represent the diversity of their designs while maintaining strong visual consistency in lighting, angle, and style. Each image was cleaned, cropped to optimal dimensions, and captioned with detailed structured descriptions covering material type, embellishment style, colour, silhouette, and setting.

Training was run on Flux Dev as the base model using ComfyUI with a custom training configuration tuned specifically for product and fashion subjects — lower learning rates to preserve base model detail rendering, combined with enough steps to firmly establish the brand’s visual signature. Multiple checkpoints were evaluated and tested against a set of reference prompts throughout the run to catch overfitting early and land on the sweet spot between brand specificity and prompt flexibility. The final LoRA activates cleanly with a single trigger token and layers naturally over any Flux pipeline.

The Results

The trained LoRA now generates on-brand Uzurii-style product imagery reliably across a wide range of prompts, settings, and compositions — without any of the manual correction that was required before. New product visuals that previously took hours of iterative prompting and editing can now be produced in minutes. The model handles different colourways, environments, and campaign styles while maintaining the core brand aesthetic throughout. Uzurii has a permanent visual production asset they own and can deploy for every future collection.

Project Details

500+

Training Images Used

10x

Faster Visual Production

1 token

Trigger to Activate Brand Style

Tools Used

Flux Dev

ComfyUI

NanaBanana

Adobe Photoshop

What Was Trained

Product design style

Material & texture rendering

Brand lighting aesthetic

Embellishment detail accuracy

Campaign visual language

The Difference

Before & After LoRA

The same prompt, run with and without the trained LoRA. Left is standard Flux output. Right is with the Uzurii LoRA active — same model, same prompt, entirely different level of brand specificity.

Before LoRA — Standard Flux output

Without LoRA — Generic output

After LoRA — Brand-consistent Uzurii output

With Uzurii LoRA — On-brand output

Sample Outputs

Generated with the LoRA

All images below were generated using the trained Uzurii LoRA — different prompts, same trigger token, consistent brand output every time.

Want your own custom LoRA?

Let’s Train Yours

Stop fighting generic AI outputs. Let’s train a custom model on your brand, product, or style — so every image you generate is on-brand from the first result.