metadata

license: apache-2.0
datasets:
  - damerajee/Llava-pretrain-small
language:
  - en
library_name: transformers
tags:
  - Vision Language Model

GPT-Vision

A very small Vision-Lanaguge Model , Like Llava and Moondream This model has THREE components combined into one

GPT2
VIT-224
Multimodality-projector

Check the github for more information GPT-Vision-Github

Inference

from transformers import AutoModelForCausalLM
from PIL import Image

model = AutoModelForCausalLM.from_pretrained("damerajee/GPT-Vision", trust_remote_code=True)

image_path = "Your_image_path"
image = Image.open(image_path)
image = image.convert('RGB')

question = "Render a clear and concise summary of the photo."
answer = model.generate(image=image,question=question,max_new_tokens=40)
print("Answer:", answer)

Limitations

A fair warning tho guys , this model is only able to generate very short response sometimes it can also repetitive generate the same tokens but even thought it will understands whats on the image

Further Fine-tuning will make this model better