Features

Custom tuning profiles

Upload 5–10 reference images of your product category — the analyzer derives a tuning profile that adjusts the cutout post-process for your subjects.

If you're cutting out a specific product category over and over (jewellery, machinery, footwear, glassware, plush toys, hardware…), the default post-processing thresholds may be slightly off for your particular subjects. Custom tuning profiles let you upload a small set of reference images, derive a profile from their statistical fingerprint, and apply that profile to every future background removal.

This is not model retraining. The model weights are unchanged. What changes is the per-pixel ΔE threshold for fringe decontamination, the erosion radius for ambiguous-pixel cleanup, and the guided-filter strength in the post-processing pipeline. All of these run client-side; nothing about your reference images leaves your browser.

Creating a profile

  1. Open the Properties sidebar in the image editor and expand the 🎯 Custom Tuning section.
  2. Type a name for the profile — e.g. "Jewellery", "Power tools", "Glassware".
  3. Click ➕ Add reference images and pick 5–10 photos that represent your typical product subject (cutouts you've done before are best — the analyzer can see what edge sharpness and partial-alpha pattern is normal).
  4. The analyzer runs immediately — no GPU required, takes a few seconds — and saves the profile to your browser's IndexedDB.
  5. The new profile is set active by default.

How profiles change the cutout

Each profile derives three overrides from the references:

  • Fringe ΔE threshold — how aggressively the cleanup pass kills background-coloured halo pixels. Soft-edge subjects (hair, fur, fabric) get a higher threshold so the edge stays soft; high-contrast subjects (logos, hardware) get a lower threshold for tighter edges.
  • Erosion radius — soft-edge subjects: 1 px (gentler); high-contrast subjects: 2 px (default); transparent subjects (glassware): 0 px (no erosion at all).
  • Guided-filter strength — soft-edge categories: 0.35 (gentler — preserves hair / fur detail); transparent: 0.25 (gentler still); other categories keep the per-quality default.

The analyzer auto-detects which category your references fit (high-contrast, soft-edge, transparent, or colour-matching-bg) based on partial-alpha fraction + edge sharpness statistics, and tunes the overrides accordingly.

Switching profiles

The Custom Tuning panel lists all your saved profiles with a radio button on each. Tick the radio to activate that profile; tick "Disable tuning (use defaults)" to fall back to the built-in defaults.

The active profile id is persisted in localStorage so it sticks across browser sessions.

Deleting a profile

Click the × next to any profile in the list. If you delete the active profile, the system reverts to defaults.

When tuning helps most

  • Catalog photography — same product type, hundreds of shots, the default may be slightly too aggressive or too gentle for your specific subjects. A tuned profile gives you 5–10 % consistency improvement across every cutout.
  • Glassware / drinkware / clear acrylic — Glass mode keeps partial alpha, but the default decontamination can still kill some of the refraction highlights. A "transparent" profile pulls back on erosion and raises fringe ΔE.
  • Hardware / tools / metallic objects — the high-contrast profile tightens the edge slightly, killing more of the background fringe that survives on hard-edge subjects.

When tuning doesn't help

  • Single-shot work — if you're processing one image, the time to build a profile outweighs the per-image quality gain.
  • Wildly varying subjects — if every image is a different category, the derived profile will be a wash.

Privacy

Reference images are analysed entirely in your browser. The profile is stored in IndexedDB on your device — never uploaded. Deleting browser data removes saved profiles.

Related