SigLIP2 Base adapted on DROID (step 20,000)

This model starts from google/siglip2-base-patch16-224 and adapts its final vision-transformer blocks and attention-pooling head to DROID robot-manipulation images. The text tower remains frozen.

import torch
from PIL import Image
from transformers import AutoModel, AutoProcessor

model_id = "YOUR_ORG/YOUR_MODEL"
processor = AutoProcessor.from_pretrained(model_id)
model = AutoModel.from_pretrained(model_id).eval()
image = Image.open("example.jpg").convert("RGB")
inputs = processor(images=image, return_tensors="pt")

with torch.inference_mode():
    result = model.get_image_features(**inputs)
    vector = result.pooler_output if hasattr(result, "pooler_output") else result
    vector = torch.nn.functional.normalize(vector.float(), dim=-1)

print(vector.shape)  # [1, 768]

Training step: 20,000.