LLaVA 1.5 7B: set up an image API in Google Colab

The original video is in Indonesian. This tutorial explains how the speaker runs LLaVA 1.5 7B in Google Colab and exposes an image-processing API through ngrok. The workflow uses Google's hosted runtime rather than the viewer's own hardware, so it is not an offline setup.

An image request returns structured text with a title, description, prompt, keywords and category. The speaker uses the resulting prompt to generate another image in Flow for a stock-image workflow. Before inference, the code converts the input image to RGB and reduces its size to shorten processing time.

The setup covers dependencies, an ngrok token and the public endpoint. Viewers learn to append /docs to the generated URL, upload an image through the Swagger interface and execute a request. The speaker then connects a custom extension to the API through its custom URL setting and /generate route. The Colab runtime must remain running, and the public URL changes between runs.

The speaker reports a 15 GB GPU allocation and roughly 12.7 GB of memory use with this model, warning that larger models can exceed the available memory. Although the title presents the API as free and unlimited, the supplied description explicitly qualifies that claim. The speaker also says the supplied notebook or custom setup involves a small charge.