Kiteet
banner
kiteet.bsky.social
Kiteet
@kiteet.bsky.social
Neo-romantic anarcho techno environmental scientist craftsman.

With AI, I have worked with painters, theatre & musical and NGOs in innovative projects.

I also write electronic music. For 💵 I work with sustainability in cryogenics. Also a gardener 🌿🌱
#gemini did a surprisingly good job of creating a composite of two old #midjourney v4 images. The more the images are iterared, the more the features change, but what is very good about Gemini (or #nanobanana actually, I think) is that the automatic segmentation Is really smart.
October 1, 2025 at 4:18 PM
#veo3 is finally an #aivideo tool that creates good and somewhat predictable result easily. I gave #geminipro some old #midjourney v4 images from way back to create composites, and Veo animated those better than I expected, really.
October 1, 2025 at 4:13 PM
I saw this tacky giant print at a shopping centre today, looks like Stable Diffusion XL + stock ESRGAN. Why not tiled ControlNet or just any tiled upscaler? Who knows, but it could have been so much more detailed 🧐 #sdxl #aiart #esrgan #digitalart #aiart
September 6, 2025 at 8:47 PM
A "spot the AI" game by the media awareness NGO Faktabaari is out. I made the images. What did I learn?

Red Panda would be terrifying if it's VAE was not weak.

Flux Dev is sometimes better than Pro Ultra. The new inpainting is bad.

SD3.5 🙄

Ideogram, Photon, etc 👎

peli.faktabaari.fi/en

#media
January 29, 2025 at 7:23 AM
In a better world, we'd have more pulp #adventure movies.

Some recent test images from #FluxDev finetunes.
January 20, 2025 at 7:52 PM
Some kind of fantasy catfox? It's interesting when LoRAs push off the capabilities of #flux1dev models and the outcomes have interesting distortions.

#animals #fantasy
January 11, 2025 at 10:33 AM
Nice slightly abstract render with a mix of my recent LoRAs

#aiart #faerytale #aiartcommunity
January 11, 2025 at 10:07 AM
LoRA mix of my own models. Nice perspective, often AI characters tend to be flat by default.

#aiart #characterconcept #ai
January 11, 2025 at 10:05 AM
More test renders with a mix of some recent Flux LORAs I've trained.

I'm really particular to organic shapes which almost but not quite resemble existing things. This fascination has been with me long before these tools.

#aiartcommunity #fantasyart #flux1dev
January 10, 2025 at 10:44 PM
Playing around with various mixes of LoRAs I've trained recently for Flux. 1 Dev. It's amazing how well it trains, considering how bland the default model is. It's almost like if they hid all of the good vectors 😅

#flux1dev #aiart #illustration
January 10, 2025 at 11:35 AM
A few mock #60s album covers from my new 60s LoRA. I've been really surprised how good Flux is with typography, especially with complex freeform logotypes.

#aiart #art
January 8, 2025 at 11:38 PM
More from my recent #60s style #lora for #flux1dev

This is a mix of three versions, all made with different optimisers, based on a quick test I prefer the mix to any of the individual models.
January 8, 2025 at 7:01 PM
I trained a LoRA on 60s silk screened posters, but the effect spread everywhere, which is just great 😁

I made five runs at 15,000 steps (n=200), two without prompts and three with them. I have the initial feeling that the ones with prompts feel slightly better, but I'll need to test more.
January 8, 2025 at 6:33 AM
I've been creating a big batch of fake images for a media awareness project. #FluxDev is the main tool obviously, and in some ways it's an amazing development over previous gen models. Still, there's interesting flaws, like tree branches which tend to grow upwards in a too orderly pattern. 1/2
#ai
December 28, 2024 at 7:15 AM
Next experiment with #stableaudioopen : I've hand curated (=my Tidal playlists which I downloaded with tidal-dl) a few thousand tracks I really like, and I'll train the base model first with those, then see how that comes out. Next I'll prepare the next version of my training set. 1/2
December 23, 2024 at 11:47 AM
I tried out the #luma Photon image model yesterday. The setup at Replicate is interesting with style and image inputs, but at least for artistic styles it feels vanilla and restricted. The inputs were abstract Disco Diffusion and SD 1.5 generations, but Photon got nowhere close in originality.
December 19, 2024 at 5:34 AM
I won the weekly challenge at #shakker, testing their DreamRender 3D model for their new Zeno base Flux model. Can't say much yet without more testing, but it seems very contrast heavy, something which I dislike about #flux in general, but of course it does allow impressive results easily.
December 14, 2024 at 10:43 AM
Today I again realised that any of my hobbies should be full time jobs 😅 I was evaluating the pile of Stable Audio Open models I made after it came out, and am leaning towards the fact that without a more robust base model finetuning it will be a bit hit and miss, and it will generalise badly.
December 13, 2024 at 11:09 PM
Interesting test I made : a logotype style transfer by training a LoRA on one image only. It's better than any of the Flux IP Adapters or the new official Inpainting model.
December 10, 2024 at 1:22 PM
#retrogaming #videogame box art model sees some progress, but with an A100 I can fit just a batch size of one with 2K training, and with a dataset of 2700 images, things are going slow...
November 25, 2024 at 6:10 AM
Found some nice places when combining some of my latest LoRAs. #flux #aiart
November 24, 2024 at 9:56 PM
I started training a #flux1dev #lora on a dataset of hand-curated 2500 retro computer game covers. So far the success has not been great, but the fails from my run with a too optimistic LR are between really interesting and hilarious 🤔 #aiart
November 23, 2024 at 9:49 PM
Playing around with custom #Flux #LoRAs I've recently made. It's curious how well Flux finetunes on art, when the base model feels lackluster with it as it is 🤔 I'm really liking some of the stuff here. I hope I had more time to take up #painting again and use these as sketches.
November 22, 2024 at 9:26 AM
I made a #flux #lora of #helsinki where I live. A rank of 128 is about the most I can fit in a 40GB A100. Flux has so far been great, but now with my set of 100 photos it surprisingly did much worse than a full #sdxl #finetune in catching the style and mood. Detail is of course better with Flux.
November 21, 2024 at 9:04 PM
I've been experimenting with running my old random #Midjourney generations through #flux to give the images a slightly more lifelike grit and also to fix some of the coherence issues with small detail. I'm not sure if MJ v4/5 were based on #stablediffusion, but anyhow I'm loving the outcomes!
November 20, 2024 at 11:53 AM