Apr 18, 2023 20:00:00

'Minigpt-4' that can chat with AI with the theme of images appears

By combining the chat AI '

Vicuna-13B ', which is said to have an accuracy comparable to ChatGPT, and ' BLIP-2 ', which can add captions to images, ' Minigpt-4 ', which allows you to talk with AI about uploaded images, Published.

Minigpt-4
https://minigpt-4.github.io/

On the page that announced Minigpt-4, you can actually touch the demonstration. First, click the part of the red frame.

Next, select the image you want to use as the theme of the conversation.

Click the 'Upload & Start Chat' button.

The upload will take some time.

When you are ready, the image and input field will be displayed again, so enter the question in the 'User' field.

An answer was displayed.

You can reset the images and conversations by clicking 'Restart'.

At the bottom of the uploaded image are the parameters for 'beam search numbers' and 'Temperature'. 'beam search numbers' seems to be a candidate for the number in

the beam search . Increasing this number tends to increase the number of answers, and even if you ask a question in Japanese, you will get an answer in English.

'Temperature' refers to the diversity of responses. However, even after fiddling with it, there was no noticeable difference.

At the time of article creation, it is said that access is concentrated, and multiple demo mirror links are provided.

Also, since the source code is open to the public, it is also possible to run it locally. Furthermore, it was said that development of a lightweight version that can be operated with one GeForce RTX 3090 is also progressing.

GitHub - Vision-CAIR/MiniGPT-4: MiniGPT-4: Enhancing Vision-language Understanding with Advanced Large Language Models

https://github.com/Vision-CAIR/MiniGPT-4