Model zoo

AI models for Card Tracking

cardstream requires exactly two models — one to find the card in the frame, and one to tell a new card from the one it is already looking at — plus an optional third that tracks the card between detections. Architecture and weights alike are Apache-2.0, so you can use them commercially without a separate licence.

`cardstream-seg-nano` Default locator

Instance segmentation — the card’s outline

Download v1

Finds the four corners of the card, so the crop sent for identification is deskewed and cut tight at the card edge instead of square with background wedges. Also stabilises the identity gate: a card tilting in the hand keeps its embedding.

Architecture: RF-DETR Seg Nano
Runtime: ONNX Runtime
Latency (CPU): ~82 ms / frame
Flag: --segmentor-model
Weights licence: Apache-2.0
Architecture licence: Apache-2.0 (Roboflow)

Version	Released	Resolution	Size	SHA-256	Download
`v1` current	2026-09	312 × 312	125 MB	`94af531c4fea1e17fcdf393f5bf1cc257a54c8a9ab1eb402e1e8b37213249fd9`	Download

`cardstream-embed-mnv2` Default gate

Embedding — the same-card identity gate

Download v1

Answers the one question that decides whether you are charged: same card, or new card? Compares crops by cosine distance between 256-dimensional descriptors. This is what makes one API call per card possible.

Architecture: MobileNetV2
Runtime: ONNX Runtime · PyTorch · LiteRT
Latency (CPU): ~5 ms / crop
Flag: --embed-model
Weights licence: Apache-2.0
Architecture licence: Apache-2.0 (Google)

Version	Released	Resolution	Size	SHA-256	Download
`v1` current	2026-09	384 × 384	9.8 MB	`f45ac9756dd621f021809f3237e5c06ce82377d4d6b1f3b8f6951804cd0369e2`	Download

`vittrack` Opt-in tracker

Visual tracking — carries the box between detections

Download 2023sep

OpenCV’s published vitTracker, served straight from the OpenCV zoo. A tracker update costs ~2.7 ms where a detect costs ~80 ms, so with --tracker-model detection drops to a slow re-sync tier and CPU real-time gets easy. Presence only — the identity gate still decides same card vs new.

Architecture: TrackerVit
Runtime: OpenCV
Latency (CPU): ~2.7 ms / frame
Flag: --tracker-model
Weights licence: Apache-2.0
Architecture licence: Apache-2.0 (OpenCV)

Version	Released	Resolution	Size	SHA-256	Download
`2023sep` current	2023-09	128 × 128	698 KB	`2990f0b7cd44d92afa48cd97db6de7be113fc1d9594fddb74e2725c10478e91d`	Download

Open source. Self-hosted. Your stream, your stack.

No platform in the middle, no seat licence: connect the Ximilar API or your own identification system, and cardstream calls it once per distinct card instead of once per frame. Whether you're breaking on Whatnot, running a Fanatics Live-style show, or streaming your own live commerce setup, run cardstream tonight on the hardware you already have.

Star on GitHub Get a Ximilar API key