SigLIP2 ViT-B/32 Vision Encoder (256)
模型来源：google/siglip2-base-patch32-256
仅裁剪保留了 Vision Encoder 部分，去除了 Text Encoder 及其他无关权重，并以 float16 精度保存为 HuggingFace SiglipVisionModel 兼容的 safetensors 格式。
Source: google/siglip2-base-patch32-256
Only the Vision Encoder is retained. The Text Encoder and all unrelated weights have been removed. Weights are saved in float16 precision as HuggingFace SiglipVisionModel-compatible safetensors format.
Model Info
| Item | Value |
|---|---|
| Architecture | SigLIP2 ViT-B/32 (fixed resolution) |
| Parameters | 94.6M |
| Hidden Size | 768 |
| Layers | 12 |
| Patch Size | 32 |
| Input Resolution | 256×256 |
| Output Tokens | 64 (8×8) |
| Weight Precision | float16 |
| File Size | ~180MB |
Usage
from transformers import SiglipVisionModel, SiglipImageProcessor
model = SiglipVisionModel.from_pretrained("path_to/siglip2-base-p32-256-ve")
processor = SiglipImageProcessor.from_pretrained("path_to/siglip2-base-p32-256-ve")
inputs = processor(images=image, return_tensors="pt")
outputs = model(**inputs) # outputs.last_hidden_state: [B, 64, 768]