From the model card
Fields below are copied from the tags and counters on the HuggingFace repository google/siglip2-base-patch16-naflex at our last fetch. They are set by the uploader, not verified by us; rows with no tag are omitted. How this page is made.
- Publisher (HF namespace)
- Pipeline tag
- zero-shot-image-classification
- Library
- Transformers
- Weight formats
- safetensors
- License tag
apache-2.0— read the license file in the repo before relying on it- Papers cited
- arXiv:2502.14786, arXiv:2303.15343, arXiv:2209.06794
- Downloads (HF counter at last fetch)
- 872,574
- Likes (HF counter at last fetch)
- 38
- Model card
- https://huggingface.co/google/siglip2-base-patch16-naflex
Use cases
- Zero-shot classification on images of varying resolutions
- Variable-resolution image embedding for retrieval pipelines
- Vision encoder for multimodal models requiring flexible input sizes
- Benchmark comparisons against fixed-resolution SigLIP models
Pros
- NaFlex handles native resolutions — no quality loss from forced resize
- Apache-2.0 license
- Base size keeps memory and compute reasonable
- Transformers-compatible with standard SigLIP2 pipeline
Cons
- NaFlex increases sequence length variability — batch padding overhead
- Base vs so400m: notably weaker on tasks requiring fine-grained visual understanding
- Requires image processor that supports flexible patch counts
- Less community documentation than fixed-resolution SigLIP variants
Tags
transformerssafetensorssiglip2zero-shot-image-classificationvisionarxiv:2502.14786arxiv:2303.15343arxiv:2209.06794license:apache-2.0endpoints_compatibleregion:us