MUSE · BAG IMAGERY · 28 OCT 2026
Bag Cutout Lab
100 bags from production's four capsules, up to four photos each, run through the cutout pipeline we'd self-host. Every image on this page was generated by it; nothing was retouched by hand.
- Masking: two open models, IS-Net (main) and u2net (second opinion), run through rembg on CPU. Their agreement decides whether a person needs to look.
- Studio cutout is the default for every bag: the real photo on the noir floor with a shadow and tier glow.
- Depth relief is an add-on for item detail: a gentle tilt built from the cutout. It looks right at moderate depth and small turns.
- Fallbacks cover the rest: a capped-size cutout for low-resolution photos, and the framed original photo when no photo cuts out cleanly.
Outcomes across the sample
Showroom · pick a bag
Each bag's final image as the app would show it. Filter by outcome or capsule.
The bench
—
Hero photou2netIS-Net
Drag to turn
No relief for this bag: it uses the framed-photo fallback, so there's no cutout to build from.
Studio cutout vs depth relief
They stack: the relief is built from the same cutout, so it's an add-on, not an alternative.
Studio cutout
- Static WebP at the app's widths (320–1280).
- Used everywhere: Showroom cards, item, vault, reveal card.
- No app work beyond pointing at a new URL.
- Always the exact bag; it stays a photo.
Depth relief
- The cutout plus a small depth map (320 px, 16-bit, about 70 KB).
- For item detail or a reveal-card tilt.
- Needs a small 3D view in the app (three.js / expo-gl or Skia).
- Holds up at moderate depth and turns; past that the invented sides show.
How it runs at ingest
A step inside the backend's image worker, once per bag, before any image reaches a phone.
- Mask every photo with both models. IS-Net and u2net through rembg, on CPU. About 3 s per photo for both.
- Score each photo and pick the hero. Agreement between the models, the mask touching the frame, fragments, a busy background and the bag's size in pixels. For Target images, request the large rendition first.
- Route the bag. Pass → studio cutout. Low-res → studio cutout at a capped size. Models disagree → studio cutout, flagged for a person in the Workbench. Nothing usable → framed original photo.
- Clean and store. Remove the old background's fringe from the edges, trim, and store the transparent cutout next to the original.
- Render. Studio images at 320–1280 px and the 16-bit depth map; the API returns their URLs with the bag.
| Step | Model | Size | Runs on |
|---|---|---|---|
| Mask | IS-Net (isnet-general-use) + u2net, via rembg | ~175 MB each | CPU, a standard Railway worker |
| Edge clean-up | pymatting foreground estimation | — | CPU |
| Depth | Depth Anything V2 Small (Apache-2.0; larger sizes are non-commercial) | ~100 MB | CPU, about 1 s |
| Optional upgrade | BiRefNet | 224 MB–1 GB | Needs over 6 GB RAM, or a GPU |
Every bag
Click a row to open it on the bench.
| Bag | Capsule | Photos tried | Model IoU | Bag px | Score | Outcome |
|---|