cstr/siglip-so400m-patch14-384-GGUF

SigLIP SoViT-400M/14 Vision Encoder 384px (GGUF)

GGUF conversion of google/siglip-so400m-patch14-384 for use with CrispEmbed.

• Architecture: SigLIP SoViT-400M/14 vision encoder (shape-optimized)
• Parameters: 428M
• Output: 1152-dimensional L2-normalized embeddings
• Input: 384x384 RGB image with SigLIP normalization (mean=0.5, std=0.5)
• Patch size: 14
• Size: ~1.6 GB
• Source: google/siglip-so400m-patch14-384

Usage

# Embed a single image
crispembed -m siglip-so400m-patch14-384 --image photo.jpg

# Batch processing
crispembed -m siglip-so400m-patch14-384 --image-dir ./photos/ --output embeddings.bin

About SigLIP SoViT-400M

SoViT-400M is a shape-optimized SigLIP variant that redistributes compute across width, depth, and MLP ratio for better efficiency. Combined with 384px input and patch size 14, it provides high-quality vision embeddings.

Notes

• All output embeddings are L2-normalized.
• This is a GGUF conversion; weights are numerically equivalent to the original HuggingFace model.