IconClip Search
Initialising…
About this model — quality & benchmarks
t2i MRR@50
—
on the IconClip Search Benchmark
nDCG@10
—
at top-10 retrieval
Size
—
downloaded once, cached
Inference
—
per-query, browser CPU
Full t2i leaderboard (Lucide slice, 1 695 icons × 120 queries)
| # | Model | MRR@50 | nDCG@10 |
|---|---|---|---|
| 1 | IconClip-ViT-L-14 (q8) | 0.505 | 0.345 |
| 2 | google/siglip-base-patch16-224 | 0.451 | 0.336 |
| 3 | laion/CLIP-ViT-L-14-laion2B | 0.399 | 0.280 |
| 4 | openai/clip-vit-large-patch14 | 0.378 | 0.239 |
| 5 | openai/clip-vit-base-patch32 | 0.298 | 0.170 |
Fine-tuning lift over base LAION CLIP-L/14: +27% MRR@50, +44% nDCG@10. See the benchmark dataset card for full methodology.
Type a query above to search across the indexed icons.
Tip: the model rewards intent over keywords —
try “throw it away”, “feeling blue”,
or “unfasten the latch”.