Next Match
Next Match finds a catalog item from a reference image plus a short edit in Turkish or English — "bunun kırmızı olanı", "çizgili olanı", "same but a hat". The test catalog is built so that the image alone (it finds look-alikes in the wrong colour) and the text alone (it finds red things of any kind) both fail; the fused query succeeds. The catalog is synthetic: procedurally drawn product icons (8 categories × 8 colours × 3 patterns).
Space: https://huggingface.co/spaces/thelamapi/Next-Match — pick a product (or upload a picture), type a change, and compare the three rankings live.
Results (1600 test queries, 192-item catalog, unseen renderings)
Recall@1 / Recall@5:
| Query group | n | Image + text (Next Match) | Image only | Text only |
|---|---|---|---|---|
| all | 1600 | 0.932 / 0.966 | 0.000 / 0.316 | 0.054 / 0.263 |
| edit=color | 648 | 0.994 / 1.000 | 0.000 / 0.358 | 0.051 / 0.218 |
| edit=pattern | 311 | 0.836 / 0.897 | 0.000 / 0.878 | 0.016 / 0.074 |
| edit=category | 322 | 1.000 / 1.000 | 0.000 / 0.000 | 0.031 / 0.183 |
| edit=color+pattern | 319 | 0.834 / 0.931 | 0.000 / 0.000 | 0.122 / 0.621 |
| lang=tr | 747 | 0.995 / 1.000 | 0.000 / 0.317 | 0.051 / 0.258 |
| lang=en | 853 | 0.878 / 0.937 | 0.000 / 0.314 | 0.057 / 0.267 |
| template=seen | 1104 | 1.000 / 1.000 | 0.000 / 0.323 | 0.057 / 0.251 |
| template=held-out | 496 | 0.782 / 0.891 | 0.000 / 0.298 | 0.048 / 0.290 |
| combo=seen | 1192 | 0.933 / 0.966 | 0.000 / 0.348 | 0.057 / 0.296 |
| combo=held-out | 408 | 0.931 / 0.966 | 0.000 / 0.221 | 0.047 / 0.167 |
template=held-out: phrasings never used in training (e.g. "show me the red one", "kırmızı olanını göster"). combo=held-out: the reference or target is one of 24 attribute combinations never rendered during training.
Method
- Image encoder: 4 conv blocks (16/32/64/64 channels) + linear → 64-d. Catalog keys are L2-normalised image features.
- Text: Turkish-aware lower-casing and longest-prefix stemming over a 61-word vocabulary (so "kırmızısı" or "olanını" still match), mean word embedding → MLP.
- Fusion: q = normalise(f_img + MLP([f_img, f_text])) — the text predicts a residual edit of the image feature.
- Baselines: image-only (image encoder + MLP head) and text-only (text encoder vs its own image keys), same architecture budget, 6000 steps, batch 128, InfoNCE with τ = 0.1.
- The Space runs all three encoders with
nlrt.json exported weights; Node parity max abs difference 1.6653345369377348e-16.
Usage
from inference import load, search
spec = load()
ids = search(spec, ref_image, "bunun kırmızı olanı") # ref_image: (48, 48, 3) uint8; ids index catalog.json
Limitations
- Toy domain: flat icons on plain backgrounds. Real product photos need a pretrained vision encoder and real catalog data.
- Vocabulary is closed: words outside it are ignored (the Space shows which words were understood).
- Uploaded photos are far outside the training distribution and are a curiosity, not a supported use.
- Trained on NVIDIA L4; profile full. Failure examples are in
evaluation.json.

