Preview images do not use other LoRAs or artist tags (Noob).
āš² Commercial use of this model is not allowed
It's okay for faces but not much else. I had a bad feeling but did it anyway because... reasons. This is the first time I train a LoRA for two characters, at least they look mostly identical... Thomas was sort of a guinea pig for me before I'd try these characters.
Negatives I use a lot: human, cropped face, another locomotive, passenger coach, goods wagon, snowplow
I also started negging car, tires after I got this...
DATASET
Most images are from episodes Love Me Tender and Emily in the Middle (the only 2 CGI episodes with Donald and Douglas having lead roles), smaller portion is from feature film Sodor's Legend of the Lost Treasure, and a few other images.
I chose to AVOID pictures showing complex formations (Donald + coach + Douglas, tenderless Douglas + Donald + coach + Douglas' tender), the dataset was deranged enough without those...
ISSUES
General deformities and coloring errors. The proportions of the engines themselves and their tenders are messed up (they end up similar to the All Engines Go style Donald toy). Not sure I could teach it what a tender is...
Can't do rear views or snowplows.
Observed with Pony: May generate a disembodied smokebox without everything behind it.
Observed with Noob: Ugly results with base Noob (assuming no style LoRAs or artist tags), their noses are typically inaccurate even in 3D style (smoothed out nose tip, nostrils), sometimes their faces turn pink (should be fixable with prompting).
Description
Finally got off my ass to TRY making this a little better, but the site drove me nuts yesterday. It refused to save my dataset changes and loaded an older one. I had to redo the tagging with Notepad... I thought the site saved the dataset the way I want it, but when I tried to import it from Re-use a Dataset it didn't work. I deleted the whole thing from the site twice or thrice until it loaded correctly. I only proceeded with the training after making sure the dataset wasn't bugging out anymore.
I figured since I won't gather 200+ images for this LoRA (too little screentime in CGI era and I don't need motion blurred characters) I would experiment with tagging.
IMAGES
Total number of images cut down to 119.
removed images of Douglas buried in snow
removed/cropped images of Douglas without his tender
removed a few other images, including a couple of near duplicates
TAGS
applied tag "3d" to whole dataset
removed instances of the twins' names and features, since they didn't work
added a bunch of tags ("train wheels", those related to facial expressions, wiped most of the old tags related to tenders and applied "tender (rail vehicle)" to some)
Trained against Pony Diffusion V6 XL on January 29-30th on-site.