Sketch-to-Product Image Model
An image-to-image model that turns rough product sketches into clean, on-brand e-commerce visuals.
The challenge
Turning early product sketches into catalog-ready imagery meant slow, expensive photo and design cycles — a bottleneck on how fast new SKUs could be tested.
Our approach
We fine-tuned an image-to-image model on the brand’s own product style so a rough sketch becomes a clean, consistent product visual. An evaluation harness scored every model version against held-out examples so quality was measured, not guessed.
- PyTorch
- Hugging Face
- MLflow
- Ray
- AWS
The build
Training and experiment tracking on MLflow, distributed runs with Ray, and a hosted inference endpoint on AWS behind a simple internal upload tool.
Turn rough product shots into clean, on-brand visuals.
Drop in a quick phone photo. Relight it, swap the background, and output store-ready imagery in seconds.
Results
Faster concept-to-visual
Usable on first pass