Kriitikon arvio on toki hänen omansa, eikä sen kanssa väitellä, mutta jos kriitikko esittää virheellistä tietoa taitelijan työskentelystä, siihen voi toki vastata. vimeo.com/1126585537/4...

Saaran ja minun yhteinen näyttely Rovaniemellä. Siinä sekoittuvat Saaran värikenttämaalaus ja oma tapani käsitellä kaupunkimaisemia, jossa tällä kertaa sekoittuvat niin Rovaniemen keskustan näkymät, Saaran maalaukset kuin myös itse galleriatila. Avajaiset 26.9. 2025 klo 18-20

Weird stares. (at ArtSibbe25 exhibition, Sipoo, Finland, today)

One piece from my Complexions series now on auction as NFT at gallery.manifold.xyz/0x1a722f5890...

I have updated my artist pages liipetti.net/visual/

Modified my urdiffusion to use EULER scheduler. Urdiffusion is for experimenting with training small models on limited private datasets overnight. Focused rather than generic aesthetics.

What if you train a diffusion model from scratch using only eight images in the training set, and train it only for half an hour.

Complexions. liipetti.net/htoyryla/sp4g/

Work in progress viewable live at liipetti.net/htoyryla/z12...

Points of view. A set of selfies run through several self developed processes. Soundtrack my own music.

My 10 years in AI art have left a trace in my Twitter/X account. This Grok summary captures the essence if it, even if the details like GPU cable can be excused as poetic license. AI art today is something quite different, I am still going my own ways.

Nyt menee mielenkiintoiseksi. Viime vuonna taidemaalarivaimoni Saara otti maalauksiensa osaksi tekoälyllä tekemiäni (ja enimmäkseen syrjään laittamiani kuvia). Nyt itse otan alunperin maisemakollaaseja kehittelevään prosessiini mukaan Saaran maalauksia, juuri niitä joissa on kuviani mukana.

A fragment. Around here. With soundtrack.

How is it possible that everything I am deeply interested in the long run turns out to be controversial. Perhaps everything is.

Some people claim that a generator must be capable of all possible visual styles to allow the artist to find his own style. I prefer to work with generators with a limited range of expression. They are difficult to work with but when you succeed you know it was worth while.

Working on my own GAN again... this is one way to ensure that there's nobody's images involved in the process but mine.

Poorly trained (or designed?) GAN used a generative tool, continued.

A GAN that fails to learn correct features from the training set as a generative process on its own right.

First train a GAN on something... say faces. Then you start a training from this model but using a different dataset. Monitor how the output changes from faces into something else, and save frequently.

You have 15 new followers. aka Return of the son of the NQGAN. Working on to revive my favorite tools from the GAN era.

Yesterday evening I felt like coding a new GAN implementation using ResNets. Checked ChatGPT for basic code for G and D. There were some errors, but it was a good exercise to find them. After figuring out good learning rates it worked quite well. Now I will start adding some additional tricks...

Adjusting the shortcuts of a U-net. There is so much one can do without any prompts.

A glimpse to my work in progress. Just now a preview of "seesteisemmin repaleista" (torn but more serenely now). A slowly changing view. liipetti.net/htoyryla/wor...

katoamassa taustaan

Stylemixer is now in Github github.com/htoyryla/sty...

Added further control at which level style features are detected, roughly on an axis from color through texture to objects and finally to deep dream like effects.

Ehdotuksia uudeksi profiilikuvaksi.

I came across a discussion in 2018, I was experimenting with feature levels in neural style. Like attenuating all but the top features. Had to try again with my recent implementation. Style model, then top features boosted, then top features attenuated.

urdiffusion experiment: model trained for just 1 hour on 64x64 crops, then run in 768x768 size, clip guided using prompt "charcoal dancing ghosts"

In less than two weeks, my exhibition Unstable Views will open in Lahti, Finland. You are welcome to visit if you happen to be close by.

My recent talk about AI art in a personal way. Almost no prompts. Own photos. Dismantling AI softwares and using the parts to build your own tools. Even tampering with the internals of AI models. vimeo.com/952877294?sh...

Experimenting with various methods for image mixing. Old fashioned style mixing makes sense visually, IP adapter works more in semantic terms.