Yet another frontier for Stable Diffusion 1.5
Even though Stable Diffusion 1.5 has proven highly effective at producing high-quality images, I didn't think it could be a viable tool to produce intricate pixel art - until I tried it out. I think it is safe to say that with the right approach, artists can leverage SD 1.5 to create stunning pixel art works. /gizmo | 2023-11-26
Check out the full gallery here
...and now the background is under control, too! This turned out better than I had hoped for! Can't wait to finish the whole video! /gizmo | 2023-10-29
| CREDITS |
Skyler (aka DJ Stosslüften): performing artist | Chris (@gizmo64k): camera, vfx & post.
...this is it! We managed to get Stable Diffusion stable enough to do animations! Judge for yourself, check out the before and after videos! Admittedly, there's still some jitter left - mainly in the background, so it is not perfect yet. But still, this increase in visual quality is really motivating! /gizmo | 2023-10-15
| CREDITS |
Skyler (aka DJ Stosslüften): performing artist | Chris (@gizmo64k): camera, vfx & post.
Using AI to index thousands of old photos and make them searchable
The benefits of deep diving into topics like generative ai are the random happy insights one gets. For example that Stable Diffusion isn't a single model but consists of multiple components such as a variational autoencoder, forward/reverse diffusion, a noise predictor, and of course a module for text conditioning, which imprints text and images into the same (latent) space and thus is the key element for text-to-image generation. Basically a word can trigger an image. And this works in the other direction, too! Long story short: I wrote a tool that scans my old photo archive and generates a brief description for each image using a locally hosted AI - no need for uploading hundreds of gigabytes to a random big tech company with questionable privacy practices - neat! /gizmo | 2023-09-13
...by coding my own Stable Diffusion inference toolkit. This way I have full control over the generation process and will be able to implement and test my own image generation concepts much more easily. Currently I use python for the back end and pygame for the front end, but this will probably change once the toolkit becomes more interactive. /gizmo | 2023-09-11