nebulette/gemma-4-vision-module

Inference

from PIL import Image
from transformers import Gemma4ImageProcessor, Gemma4VisionModel

if __name__ == '__main__':
  device = 'cuda'
  vision = Gemma4VisionModel.from_pretrained('nebulette/gemma-4-vision-module').to(device)
  processor = Gemma4ImageProcessor.from_pretrained('nebulette/gemma-4-vision-module')
  inputs = processor.preprocess(Image.open('path/to/image.png'), return_tensors='pt').to(device)
  outputs = vision.forward(pixel_values=inputs.pixel_values, pixel_position_ids=inputs.image_position_ids)
  hidden_state = outputs.last_hidden_state