Snacks vision model: counterfeit detection

What it covers

The product types the model is trained on. Anything outside this list is handled by a neighbouring model.

In scope

Chips and crispsBiscuits and cookiesChocolate and candy barsNuts and dried fruitSavory and protein snacks

Not covered

Packaged-foods for mealsSoft-drinksDietary-supplementsBaby-formulaFresh unpackaged produce

What it reads

The model reads snacks through a structured taxonomy. Below are its top-level feature families. Each one breaks down into detailed parts, attributes and possible values, and some of our taxonomies go beyond 200 labels. The full taxonomy is shared during a pilot.

01Container geometry
02Closure and seal construction
03Panel and shoulder profile
04Label architecture
05Surface finish
+Detailed parts, attributes and values within each family

Output format

Once live, the model returns the same record as every Klaimed model: the design similarity, a score for each feature, and the evidence image.

POST /v1/detect/snacksPreview

{
  "model": "detect-snacks",
  "listing": "site_a/8812-4410",
  "similarity": 0.94,
  "attributes": {
    "container_geometry": 0.93,
    "closure_and_seal_constructio": 0.94,
    "panel_and_shoulder_profile": 0.95,
    "label_architecture": 0.96,
    "surface_finish": 0.49
  },
  "evidence": "ev_9f3a.png"
}

How teams use it

IP firmsMonitor snacks copies for your clients, under your brand Design search firmsSearch snacks designs across brand websites and registries Brand-protection platformsAdd snacks recognition to your pipeline by API

Capabilities

Fine-grained attribute understanding: Reads snacks the way a category expert would: overall shape and proportion, the geometry of each component, material and surface character, and the specific construction features that define one design against another.
Design-level similarity retrieval: Given a reference product, returns the closest matches from a monitored set, ranked by design similarity rather than by colour, scene or filename.
Feature-by-feature match explanation: Every match comes with its reasoning: which features agree between the reference and the candidate, which diverge, and how confident each comparison is.
Stable across how the product was photographed: The same design retrieves consistently whether it was shot as a studio packshot, a seller's phone photo or an in-use image, and holds up under changes of angle, lighting, background and partial occlusion.

Questions

Can the snacks model find counterfeits without a logo or brand name?

Yes. It reads snacks by design, through features such as container geometry, closure and seal construction and panel and shoulder profile. A listing with no logo, no brand name and a different photo still matches its reference when the design is the same.

Which product types does the snacks model cover?

In scope: Chips and crisps, Biscuits and cookies, Chocolate and candy bars, Nuts and dried fruit and Savory and protein snacks. Not covered: Packaged-foods for meals, Soft-drinks, Dietary-supplements, Baby-formula and Fresh unpackaged produce.

When will it be available?

It is in development. Early-access partners help shape the model on their own products. Tell us the product type and your volume and we'll share the timeline.

What images does it accept?

Studio packshots, seller phone photos and in-use images. Results stay consistent across angle, lighting, background and partial occlusion.

Is the attribute taxonomy public?

No. The taxonomy and training data are proprietary. We walk you through it during a pilot, and can add features and retrain the model on request.

How do I access it?

Request early access. Once it is live, it runs through the same Klaimed API and interface as every other model.

Snacks vision model