DeepSeek Janus Pro: local setup and ChatGPT comparison

Uploaded on January 30, 2025, this tutorial covers DeepSeek Janus Pro's 1B and 7B models. The demonstrated setup uses Windows, Python 3.8 or newer, and a PyTorch installation selected for CUDA 12.1. Its dependency workaround and installation steps reflect that upload period.

The speaker first tests the Hugging Face demo, then explains how to run models locally through the repository's Gradio interface. Image understanding examples include interpreting a meme, converting a photographed equation into LaTeX, reading handwriting and identifying people in a television screenshot. Results vary: the handwriting response guesses an unfinished word, and Janus misidentifies the characters. ChatGPT gives a more accurate answer to that screenshot in the speaker's comparison. He finds image generation less convincing and reports output at 384 by 384 pixels.

The local setup covers downloading the repository, installing dependencies, switching between model sizes and disabling Gradio's public sharing link. The speaker encounters a SentencePiece installation error and changes its dependency version as a workaround. Model downloads precede offline use. A CPU image-generation attempt errors after roughly seven minutes; the subsequent GPU example produces an image, with poor text rendering.

GPU testing uses a 12 GB RTX 3080 Ti with the 1B model. The speaker reports an image-question response in roughly two seconds, versus about 30 seconds on CPU, though identification errors remain. He suggests around 16 GB of VRAM for 7B; he does not demonstrate that model on his GPU.