@htoyryla

Visual artist, engineer, medievalist. AI ART my own way since 2015.

 Helsinki, Finland
liipetti.net/visual
Joined September 2012
  • Tweets3.8K
  • Following254
  • Followers841
2.1KMedia
Ahmed has made a great tool which uses xurdif in more ways than one can imagine. Train models, generate images, bend and mix models and so on. Compatible with my latest xurdif but it is more than just an UI for xurdif. Works on Windows, Mac and Linux. github.com/abuzreq/kiln

GitHub - abuzreq/Kiln: Train small diffusion models on your own pictures, then create with them or...

Train small diffusion models on your own pictures, then create with them or take them apart. - abuzreq/Kiln

github.com
3
88
Two urdiffusion studies. Interpolating between random faces in init noise space, but changing target before each endpoint. Small skip to make the model dominate over realism. Attenuated init noise to further reduce detail.
1
54
Yesterday's experiment continued. Visions of history no. 781. On soundtrack electric piano and acoustic Andean flute.
2
71
Took photos of leaves on the ground... there are plenty of them right now. Then trained a tiny diffusion model for two hours to get this look. No prompting, just interpolation in the initial noise space.
2
4
168
With urdiffusion, you are not training to images that look like the training images. You are experimenting what look you will get with your own small datasets. Which can be just anything, and as you work further, you accmuulate more datasets too.
2
160
My window 50 years ago.
37
Gradually changing portrait. Completely homemade small diffusion model used to transform the visual appearance of a series of photos and to produce a gradually changing image flow.
1
3
196
...and in fact, also this model was trained using this dataset...possible as the model is not trained for primarily semantic but visual features.
22
Revisiting a dataset from the GAN age in 2018, now with my #urdiffusion.
1
38
I mean this experiment...
22
Here training #urdiffusion from scratch with a new dataset. First 90 minutes. Big difference from the GANs 8 years ago.
3
237
#urdiffusion model trained from scratch for 1 hour, then using init images, interpolating at noise level (like latent interpolation but at pixel level). This is my kind of AI art. Simple models, small datasets, plain diffusion, no prompts, custom code.
3
1
7
448
For instance... if I want to get rid of the brighter red spots, I cannot use prompt "avoid bright red". I have to think in terms of image processing, how to avoid those pixel ranges in my process.
1
39
And as I have built the process, I can tamper with any phase in the process, even down to the internals of the diffusion model.
23
Source: Tomi Slotte Dufva: Entanglements in AI Art acris.aalto.fi/ws/portalfile…
1
1
4
149
I like how Tomi condensed my approach: "AI as a material which takes part in the art-making process". Especially the wording "art-making" instead of "artistic". In a way, material has always participated in art.making, its role increasing over time.
20
A frame from a stream of changing views of my childhood hometown. Photos processed with #urdiffusion, then made into a stream of partial crossfades, scratched and descratched and what else...
52
My new toy from yesterday may sound like meh... segment image and process separately, so what. But this is urdiffusion, if you start experimenting, you never know where you will end.
1
3
62
"Defending my thesis" 2-way urdiffusion (new web app using separate model and settings for back/foreground)
2
95
1
1
91
18
Load more