siglip2-large-lora-v3

LoRA fine-tune of google/siglip2-large-patch16-384 for material/surface visual-similarity retrieval (room crop → catalog product), trained with Supervised Contrastive loss on a combined v1-mined + v3-synthetic dataset.

Retrieval eval (crop → full catalog gallery)

n	recall@1	recall@5	recall@10
845	0.303	0.5136	0.6154

Per-category / per-slot

stratum	n	r@1	r@5	r@10
slot:wall	111	0.0811	0.2793	0.3604
slot:floors	87	0.046	0.2184	0.3678
slot:ceiling	78	0.0641	0.2692	0.4231
cat:Fabric & Textile	47	0.6596	0.8936	0.9787
slot:table	42	0.0714	0.3333	0.4762
slot:rug	41	0.122	0.3902	0.6098
cat:Wallpaper & Wallcovering	36	0.6944	0.8611	0.9444
slot:cabinetry	36	0.0833	0.3333	0.5278
cat:Handle / Knob	35	0.4857	0.7714	0.8571
slot:sofa	34	0.1176	0.3824	0.4412
cat:Tile	33	0.697	0.8788	0.8788
slot:chair	28	0.0357	0.2857	0.3929
cat:Stone & Masonry	27	0.8148	1.0	1.0
cat:Flooring	26	0.7308	0.9615	1.0
cat:Paint	21	0.619	0.8095	0.8095
cat:Carpet	18	1.0	1.0	1.0
slot:cladding	17	0.0588	0.1765	0.2353
slot:worktop	16	0.0	0.25	0.5
cat:Wood	13	0.6154	0.9231	0.9231
cat:Engineered Surface	13	0.7692	0.9231	0.9231
slot:rod_bar	13	0.0769	0.3846	0.6154
slot:curtain	12	0.0833	0.3333	0.4167
cat:Glass	12	0.75	0.9167	1.0
cat:Laminate	10	0.1	0.3	0.5
cat:Leather	10	0.9	1.0	1.0
cat:Metal	9	0.3333	0.7778	1.0
cat:Faucet / Tap	8	0.75	1.0	1.0
cat:Plastic	6	0.8333	0.8333	0.8333
slot:moulding	6	0.0	0.0	0.0

Training

• epochs run: 10
• final train_loss: 0.0900
• LoRA r=16, α=32, dropout=0.1; SupCon τ=0.07; batch 16; lr 1e-4 cosine

Load

from peft import PeftModel; from transformers import AutoModel
base = AutoModel.from_pretrained('google/siglip2-large-patch16-384')
model = PeftModel.from_pretrained(base, 'subhrokomol/siglip2-large-lora-v3')