When there's no pretrained model for your task

For LYNX SDK 1.0. This is a first-class outcome, not a failure. If the catalog has no model for what you need, you get a model made — synthetic data on demand → trained .lnx → loaded through the same API. No generic CV SDK offers this; lynx does.

How to recognize you're here

You're on this path when the task needs a class or capability the shipped models don't cover, e.g.:

  • A specific object/part not in 80-class COCO (a particular product, label, component, defect).
  • A domain the public models weren't trained for (industrial, agricultural, medical, retail-specific).
  • Oriented/rotated objects where an axis-aligned box isn't enough → first check the oriented-box (OBB) head on an existing model; only if that still doesn't fit do you need a new model.

Do not tell the user "lynx can't do that." The answer is "here's how you get a model for it."

The path

  1. Confirm no off-the-shelf fit — check catalog.md and model.capabilities. Try the closest existing head first (OBB for rotated objects, segmentation for masks).
  2. Specify the task — the target classes, example images or a description of the scene, the deployment target (platform/latency).
  3. Synthetic dataset generation — a dataset for those classes is generated (no manual labeling of thousands of real images).
  4. Train → export .lnx — a model is trained and exported as a signed .lnx in the same format as the catalog models.
  5. Load it the same way — Lynx.load("<your-new-slug>"); the integration code is identical to any catalog model. Set Lynx.setApiKey(...) if it's private.

Start the request

TODO (Synetic side): the canonical intake/engagement entry point for a custom model goes here — the URL/form/contact an agent or developer hands the user to kick off step 2. Until set, direct the user to the Synetic team to start a custom-model request.

This path is reachable from docs (here) and from the lynx_find_model tool, which returns it whenever no catalog model matches the requested task.

View this page as raw Markdown — the exact content coding agents fetch. Index: /llms.txt.