Andrea Botez is a Canadian chess player and streamer, the younger sister of Alexandra Botez. I've trained her TI for 2500 steps on a dataset of around 25 images. After many tries and tests, I've decided, for the time being, to settle on step 1850, which is probably a bit overcooked already but still is producing some great results for me.
If it feels like you're getting oversaturated results, try reducing the embedding weight to something like 0,9 (or even 0.95). That should help quite a bit. Enjoy!
As with her sister, I've applied my new training method to my dataset of Andrea. The new TI has been trained for 6550 steps and seems less saturated and probably more accurate than the previous version, so I'm uploading it in case you wanna try it. Have fun!
By the way, regarding prompt-building, I suggest you start the prompt with "a photo of", "a picture of", etc., and then you add "4ndreabotez". It seems to be much more precise that way. I've seen results that didn't have much to do with the subject when I started the prompt directly with "4ndreabotez". This is probably a consequence of how I train my models and the prompt template and filewords I use, but it seems unavoidable.
Go ahead and upload yours!
Your query returned no results β please try removing some filters or trying a different term.