Today I was playing with a new 2D-image-to-3D AI models called "DreamGaussian". I gave it a picture of me and asked the AI to translate it to 3D.

Fun to play with, but the quality and detail of the output is still low. I'm sure it'll get better soon.

Was playing around with ComfyUI today.

It's a tool to let you generate AI images+videos (e.g. via Stable Diffusion) on your own machine.

Harder to use than Pika or RunwayML, but easier than automatic1111.

For full control, or if have your own GPU, try this.

Today I learned that your website can have a `og:video` <meta> tag.

iMessage will render the video in the preview.

Trying out HeyGen today - please meet AI video Gabor.

I trained it on a webcam recording, and then typed in what it should say on my behalf.

Feels maybe a tiny bit robotic, but still incredible.

I'm taking an RL (reinforcement learning) class and one of the tasks is to re-implement the classic 2015 DeepMind Atari paper.

Here's the same deep Q-learning network we built for the class, trained with Pong + Space Invaders. 1.7M parameters. Tiny tiny by today's standards.

I spent this week playing with generative AI video tools.

Pika and Runway both take still images and animate them. So I made this mini sci-fi trailer trying to use both.

Pika is pretty good, but at this point it's Discord-only which I find hard to use.

Runway has much better UI, but the results can be awkward when you don't guide it with their magic brush feature.

Playing around with RunwayML's Magic Brush feature.

Here's a still image of the Simplon Pass hospice I took on our summer vacation in Switzerland, animated into a dramatic 8s video.