See our collection for all versions of OWL-ViT.

Run OWL-ViT with Keras 3: JAX, PyTorch, or TensorFlow

zeromodels/owlvit-base-patch16

Paper: Simple Open-Vocabulary Object Detection with Vision Transformers (arXiv:2205.06230) · HF Papers

OWL-ViT detects objects described by free text, with no fixed class list. It starts from a CLIP-style vision and text encoder, then drops CLIP's pooling and attaches a lightweight box head to every patch token. Each patch becomes a detection candidate, scored by cosine similarity against your text queries.

For more details on the model, please go to Google's original model card.

Pure-Keras 3 conversion of google/owlvit-base-patch16 for zeromodels. One implementation runs unmodified on TensorFlow / Torch / JAX.

This is an open-vocabulary object detection checkpoint (OwlViTDetect): pass free-text prompts at inference time.

✨ Quick start

import os
os.environ["KERAS_BACKEND"] = "torch"  # or "jax" / "tensorflow"

from PIL import Image
from zeromodels.models.owlvit import (
    OwlViTDetect,
    OwlViTProcessor,
    OwlViTImageProcessor,
)

model = OwlViTDetect.from_weights("zeromodels/owlvit-base-patch16")
processor = OwlViTProcessor.from_weights("zeromodels/owlvit-base-patch16")
image_processor = OwlViTImageProcessor.from_weights("zeromodels/owlvit-base-patch16")

image = Image.open("your_image.jpg").convert("RGB")
prompts = ["a photo of a mug", "a photo of a knife"]
inputs = processor(text=[prompts], images=image)
output = model(
    {
        "input_ids": inputs["input_ids"],
        "pixel_values": inputs["pixel_values"],
    }
)
results = image_processor.post_process_object_detection(
    output,
    threshold=0.1,
    target_sizes=[(image.height, image.width)],
    text_labels=[prompts],
)[0]
for score, name, box in zip(
    results["scores"], results["text_labels"], results["boxes"]
):
    print(f"{name}: {float(score):.3f} {box}")

Load any OWL-ViT / OWLv2 variant the same way with from_weights("zeromodels/<variant>") (use OwlViTDetect for this repo):

Tips

• Set KERAS_BACKEND before importing Keras / zeromodels.
• Prefer Processor.from_weights(...) so image size matches the variant.
• Open-vocab thresholds are often much lower than closed-set detectors (try 0.1).
• OWLv2 pads to square before resize; pass the original (height, width) as target_sizes carefully (see the OWLv2 docs for the padding trap).
• See OWL-ViT docs and Loading Weights.
• Community / upstream safetensors still work via the hf: prefix, e.g. OwlViTDetect.from_weights("hf:google/owlvit-base-patch16").

Special Thanks

A huge thank you to the Google OWL-ViT authors for creating and releasing these models.

License: Apache 2.0.