🚀 Image understanding playground, engined with DeepSeek Janus Pro 1B + FastAPI

• This repository is the source code of website, which supports AI image reasonning locally as a server
• Website is built by using FastAPI
• Local AI engine is Deepseek Janus Pro 1B, the Multimodal Understanding and Generation Models
• Below is a screenshot of demo

Prerequisites:

• Python 3.10, anaconda
• PyTorch >=2.0.1
• CUDA, GPU suggests no lower than Nvidia RTX3090
• Ubuntu 22.04/20.04

Installation:

• Clone this repository and enter to this repository name

• Create conda environment (optional), e.g. name of web

• Install dependencies in your prefered environment:

  cd img_chat_deepseek
  conda create --name web python=3.10
  conda activate web
  pip install -e .  
  

Copy weights of Janus

• URL: https://huggingface.co/deepseek-ai/Janus-Pro-1B
• Copy all files under "File" tab to folder of "Janus-Pro-1B"
• If you want to use 7B version, download it and revise the associated code in img_ana.py

Implement

• Enter to this repository folder
• Under img_chat folder, start server:

  uvicorn app.main:app --reload

• visit 127.0.0.1:8000 in browser

Acknowledgement

• The AI model is forked on deepseek-ai Janus.
• Janus' source code is under the folder of "Janus"