NSS Background Remover
Custom tuning: optimize the pipeline for your kind of image
Upload reference images to tune the removal pipeline for a specific category of image — your products, your style — so cutouts get more consistent across a catalog, all on-device.
When you process the same kind of image over and over — a catalog of similar products, a consistent style of shot — a general-purpose pipeline can be tuned to do better on your specific case. NSS Background Remover supports custom tuning: an active tuning profile that flows into the inference worker, driven by reference images you provide.
This tutorial covers how custom tuning works, when it helps, and how it keeps cutouts more consistent across a batch — all while staying on-device.
- 1
1. Decide when tuning is worth it
Custom tuning pays off when you have a consistent category of image — the same kind of product, the same lighting, the same style — and want cutouts to behave consistently across all of them. For one-off images, the default pipeline is already strong and tuning is unnecessary.
Think of tuning as something you set up once for a repeated job, not for every image.
- Best for a consistent category processed repeatedly.
- Unnecessary for one-off images.
- 2
2. Provide reference images
Upload reference images that represent your category. These shape a tuning profile that optimizes the pipeline for that kind of image — so the cutouts on your actual catalog reflect the characteristics of the references.
As with everything in the tool, the references are processed on your device; tuning does not send your images anywhere.
- Reference images define the tuning profile.
- Processed on-device, like the rest of the tool.
- 3
3. Let the profile flow into inference
The active tuning profile flows into the inference worker, so subsequent removals on your category benefit from it. The result is more consistent behavior across a batch — fewer one-off surprises on images that are all basically the same kind.
Consistency across a catalog is exactly what makes a set of listings look like one coherent brand rather than a pile of unrelated photos.
- The tuning profile feeds the inference worker.
- More consistent cutouts across a similar batch.
Tune once, batch consistently
Pair custom tuning with batch processing: set the profile from references that represent your catalog, then run the batch so the whole set is cut out consistently. The combination is what turns a pile of similar product shots into a uniform, marketplace-ready set.