moondream2 is a small vision language model designed to run efficiently on edge devices. Check out the GitHub repository for details, or try it out on the Hugging Face Space!
Benchmarks
| Release | VQAv2 | GQA | TextVQA | DocVQA | TallyQA(simple/full) | POPE(rand/pop/adv) |
|---|---|---|---|---|---|---|
| 2024-07-23 (latest) | 79.4 | 64.9 | 60.2 | 61.9 | 82.0 / 76.8 | 91.3 / 89.7 / 86.9 |
| 2024-05-20 | 79.4 | 63.1 | 57.2 | 30.5 | 82.1 / 76.6 | 91.5 / 89.6 / 86.2 |
| 2024-05-08 | 79.0 | 62.7 | 53.1 | 30.5 | 81.6 / 76.1 | 90.6 / 88.3 / 85.0 |
| 2024-04-02 | 77.7 | 61.7 | 49.7 | 24.3 | 80.1 / 74.2 | - |
| 2024-03-13 | 76.8 | 60.6 | 46.4 | 22.2 | 79.6 / 73.3 | - |
| 2024-03-06 | 75.4 | 59.8 | 43.1 | 20.9 | 79.5 / 73.2 | - |
| 2024-03-04 | 74.2 | 58.5 | 36.4 | - | - | - |
Usage
pip install transformers einops
from transformers import AutoModelForCausalLM, AutoTokenizer
from PIL import Image
model_id = "vikhyatk/moondream2"
revision = "2024-07-23"
model = AutoModelForCausalLM.from_pretrained(
model_id, trust_remote_code=True, revision=revision
)
tokenizer = AutoTokenizer.from_pretrained(model_id, revision=revision)
image = Image.open('<IMAGE_PATH>')
enc_image = model.encode_image(image)
print(model.answer_question(enc_image, "Describe this image.", tokenizer))
The model is updated regularly, so we recommend pinning the model version to a specific release as shown above.