Krea2
Showcase images :
Turbo model
768x1152
Euler / simple / 8 steps / CFG 1
Lora strength 1-1.5
Klein 9B
Showcase images :
Distilled model
768x1152
Euler / normal / 4 steps / CFG 1
Lora strength of 1
Qwen
I keep experimenting with qwen lora training, so here a lora trained on screencaps from movies "Mon oncle" and "Playtime" directed by Jacques Tati.
The dataset was around 100 images, starting the caption with "tatitati movie still of"
Description
FAQ
Comments (2)
I'm astounded by some of these results. It's crazy how many details it captures from the non-lora model, or that are included in your other 2D art styles. It translates it all into something coherently like this film style and technology/film art of the time, like with the fake animals/dioramas. So good.
So you used 100 images. How high resolution/clear are your images? Do you use blu-ray sources? Does that kind of source matter? You train so many loras. Do you tweak settings to be different between 2D art styles and film styles? Do you tweak anything at all? What do you train in? Can you share settings?
I'm just so impressed with this lora, and it's closest to the one I've built a dataset for, late Showa era tokusatsu TV show. Thanks for any insights you can share!