ModelListadosExterno

Siglip2 Large LoRA v1

LoRA fine-tune of google/siglip2-large-patch16-384 for material/surface visual-similarity retrieval. Trained with Supervised Contrastive loss on catalog↔in-room segment pairs.

Fontehuggingface.co/subhrokomol/siglip2-large-lora-v1/tree/e22a7bb6328b09cb75730de50abf8f886ad25873

TipoOtherModelo basegoogle/siglip2-large-patch16-384VersãoAdapter modelBnota

Fonte do modelo

Tipo: Other
Tamanho do arquivo: 12,2 MB
Modelo base: google/siglip2-large-patch16-384
Versão: vAdapter model
Atualizado: 22 de out. de 2026
Fonte: Hugging Face

Descrição da fonte

SigLIP2-Large LoRA v1

LoRA fine-tune of google/siglip2-large-patch16-384 for material/surface visual-similarity retrieval. Trained with Supervised Contrastive loss on catalog↔in-room segment pairs.

Best validation metrics

recall@1	recall@5	recall@10
0.2795	0.5486	0.6562

Training configuration

Base model: google/siglip2-large-patch16-384 (vision encoder)
Adapter: LoRA, α=32, r=16, dropout=0.1
Target modules: q_proj, k_proj, v_proj, out_proj (attention)
Loss: Supervised Contrastive (SupCon)
Temperature: 0.07
Batch size: 32 (×1 grad accum)
Images per class: 2
Optimizer: AdamW, lr=1e-4, weight_decay=0.01
Schedule: cosine, 10% warmup
Epochs trained: 4
Final train loss: 0.2828

Per-epoch metrics

epoch	train_loss	recall@1	recall@5	recall@10	intra	inter	sep
1	0.4293	0.2674	0.5486	0.6562	0.7547	0.2508	3.0096
2	0.3911	0.2795	0.5365	0.6562	0.7380	0.2304	3.2032
3	0.3208	0.2604	0.5312	0.6562	0.7273	0.1999	3.6388
4	0.2828	0.2691	0.5434	0.6562	0.7238	0.1936	3.7391

intra — mean cosine similarity for same-class pairs
inter — mean cosine similarity for cross-class pairs
sep — intra / inter (higher = better class separation)

Companion dataset

subhrokomol/siglip2-large-lora-v1-dataset

Load

from peft import PeftModel
from transformers import AutoModel, AutoProcessor

base = AutoModel.from_pretrained("google/siglip2-large-patch16-384")
processor = AutoProcessor.from_pretrained("google/siglip2-large-patch16-384")
model = PeftModel.from_pretrained(base, "subhrokomol/siglip2-large-lora-v1")
model.eval()

# image embedding (1024-dim)
import torch
from PIL import Image
img = Image.open("your_image.jpg").convert("RGB")
inputs = processor(images=img, return_tensors="pt")
with torch.no_grad():
    embed = model.get_image_features(**inputs)
embed = embed / embed.norm(dim=-1, keepdim=True)  # L2-normalize

Intended use

Visual-similarity retrieval over materials, surfaces, and textures — e.g. given a crop from an interior photo, retrieve catalog products with similar visual appearance. Cross-domain (in-the-wild photo ↔ catalog studio photo) is the primary use case.

License

Apache 2.0.

Fontes

1 fonte

Hugging FacePrimárioVersão e22a7bb6328b09cb75730de50abf8f886ad25873

huggingface.co/subhrokomol/siglip2-large-lora-v1/tree/e22a7bb6328b09cb75730de50abf8f886ad25873

Verificado 22 de out.

Artefatos de modelo

1 artefato

adapter_model.safetensors

safetensors · 12,2 MB · SHA-256 afde10b09175…3faf · Hugging Face

Download file

Trechos de fonte

2 trechos

Readme5 de ago.huggingface.co/subhrokomol/siglip2-large-lora-v1

--- license: apache-2.0 base_model: google/siglip2-large-patch16-384 library_name: peft tags: - siglip - siglip2 - lora - peft - embedding - visual-similarity - material-recognition - contrastive-learning - supervised-contrastive pipeline_tag: image-feature-extraction --- # SigLIP2-Large LoRA v1 LoRA fine-tune of [`google/siglip2-large-patch16-384`](https://huggingface.co/google/siglip2-large-patch16-384) for material/surface visual-similarity retrieval. Trained with Supervised Contrastive loss on catalog↔in-room segment pairs. ## Best validation metrics | recall@1 | recall@5 | recall@10 | |---:|---:|---:| | 0.2795 | 0.5486 | 0.6562 | ## Training configuration - **Base model**: `google/siglip2-large-patch16-384` (vision encoder) - **Adapter**: LoRA, α=32, r=16, dropout=0.1 - **Target modules**: `q_proj`, `k_proj`, `v_proj`, `out_proj` (attention) - **Loss**: Supervised Contrastive (SupCon) - **Temperature**: 0.07 - **Batch size**: 32 (×1 grad accum) - **Images per class**: 2 - **Optimizer**: AdamW, lr=1e-4, weight_decay=0.01 - **Schedule**: cosine, 10% warmup - **Epochs trained**: 4 - **Final train loss**: 0.2828 ## Per-epoch metrics | epoch | train_loss | recall@1 | recall@5 | recall@10 | intra | inter | sep | |------:|-----------:|---------:|---------:|----------:|------:|------:|----:| | 1 | 0.4293 | 0.2674 | 0.5486 | 0.6562 | 0.7547 | 0.2508 | 3.0096 | | 2 | 0.3911 | 0.2795 | 0.5365 | 0.6562 | 0.7380 | 0.2304 | 3.2032 | | 3 | 0.3208 | 0.2604 | 0.5312 | 0.6562 | 0.7273 | 0.1999 | 3.6388 | | 4 | 0.2828 | 0.2691 | 0.5434 | 0.6562 | 0.7238 | 0.1936 | 3.7391 | - `intra` — mean cosine similarity for same-class pairs - `inter` — mean cosine similarity for cross-class pairs - `sep` — intra / inter (higher = better class separation) ## Companion dataset [`subhrokomol/siglip2-large-lora-v1-dataset`](https://huggingface.co/datasets/subhrokomol/siglip2-large-lora-v1-dataset) ## Load ```python from peft import PeftModel from transformers import AutoModel, AutoProcessor base = AutoModel.from_pretrained("google/siglip2-large-patch16-384") processor = AutoProcessor.from_pretrained("google/siglip2-large-patch16-384") model = PeftModel.from_pretrained(base, "subhrokomol/siglip2-large-lora-v1") model.eval() # image embedding (1024-dim) import torch from PIL import Image img = Image.open("your_image.jpg").convert("RGB") inputs = processor(images=img, retur...

Note5 de ago.huggingface.co/subhrokomol/siglip2-large-lora-v1

Source context: 3 downloads · 0 likes · Pipeline image-feature-extraction · Library peft · Repo subhrokomol/siglip2-large-lora-v1