Dec 21, 2023 20:00:00

'Image Telephone' shows you how the image changes when you repeat the process of converting a document output by chat AI into an image using image generation AI over and over again like a message game

The image generated by AI is created by repeating the process of ``inputting an image into chat AI and adding a caption, converting it into an image using image generation AI

DALL-E , inputting it into chat AI again...'' ' Image Telephone ' allows you to observe how things change. The author, elijahbenizzy , explains , ``When I tried it myself, the results were very interesting, so I decided to make it into an app so that anyone can easily experience it.''

Image Telephone · Streamlit
https://image-telephone.streamlit.app/

Image Telephone is an application to find out what happens when you play a message game with GPT-4 (the large-scale language model that is the basis of ChatGPT) and DALL-E 3. First, let GPT-4 create a caption for the input image, then input the caption to DALL-E 3 to generate the image. Image Telephone allows you to repeat this process over and over again and see how the image changes. Image Telephone does not execute all steps in real time, but only hosts pre-executed content, and the reason for this is stated that ``execution is time-consuming and expensive.'' Masu.

On the Image Telephone screen, a cafe latte is displayed first.

As you scroll the screen, you can see how the output image changes as a result of the message game between AIs. The number written under the image shows how many times the message game loop has been executed, and '0' in the upper left is the original image used at the beginning of the message game.

The first few times, an image depicting the liquid level of the cafe latte was output, but as the number of message games increases, the liquid level is no longer drawn, and only the surrounding pots and cups increase.

The number of pots has increased and the cafe latte element is no longer there.

In the end, for some reason, it ended up being displayed in a place like a museum.

The first image input can be changed from the pull-down menu at the top of the screen.

In addition to latte, various samples are available such as Christmas house (christmas_house), waterfall (waterfall), Windows XP (windows_xp), and Golden Gate Bridge (golden_gate_bridge).

Also, by moving the slider under the pull-down menu left or right, you can display the 'image output as a result of the XX message game'.

Furthermore, click 'Caption' below the image.

Then, you can check the caption that is output when the above image is input to GPT-4.

Please note that Image Telephone is built using Hamilton , a declarative UI framework for building clean, self-documenting, and extensible data pipelines, which takes data flows from the Hamilton Dataflow Hub and creates your own. Generating image.