Gene Kogan
@genekogan

programmer, primate @ml4a_ @abraham_ai_ https://t.co/BOIBaSFktS




Gene Kogan    @genekogan
Sometimes if I want advice about something, I start drafting a long message to someone about it. Writing it down forces me to think so clearly that I end up answering my own question, and never end up sending it. The lesson is: write! Writing is an upgraded version of thinking.

Gene Kogan    @genekogan
A/B testing the new CLIP weights released yesterday. So far I can't say I prefer one to the other, but it's nice to have another hyperparameter to toy with. Left: ViT-B/32 (the usual), Right: ViT-B/16 (new)

Gene Kogan    @genekogan
Averaging CLIP embeddings over many permutations of the form "a photograph of X", "X trending on artstation", "X rendered in Blender/Unity," etc produces fairly dramatic improvements in some cases. e.g. "Jabba the Hutt smoking a cigar" Left: single text, Right: avg many texts

Gene Kogan    @genekogan
A tip for the CLIP ninjas. Multi-scale rendering converges faster and gives better quality. e.g. optimize for a while at 128x128, then upsample to 256x256 and run some more, then again at 512x512. As a bonus, you can afford bigger batches at low-res and can pick the best ones.

Gene Kogan    @genekogan
More recently, I've been making progress in applying the looping technique to text-to-image pipelines. I'm hoping to release something like this soon, once I get it to where I like. https://t.co/X5vzirjIaU

Gene Kogan    @genekogan
Whenever I try the "free desktop wallpaper" trick with CLIP + VQGAN, I get a stock photography watermark

Gene Kogan    @genekogan
Editing real images with CLIP: a picture of the moon -> "A lush, tropical planet full of exotic lifeforms" Natural language is the future of HCI.

Gene Kogan    @genekogan
While I enjoy random ears and noses strewn about the canvas as much as the next person, it is nice to get coherent faces once in a while.











 








Gene Kogan

John Carlos Baez

Facebook AI

Mark O. Riedl

DeepMind

Daniel Silva