So here's a character who you meet briefly in Kotor 2. So far I don't think a lora's been made for this character so I decided to train one. I fed some images into GPT for a more modern look for this character beyond the dated graphics in order to make a training set. Since this model was trained on screen caps that all had the same background it might be necessary to turn down the character's lora strength if you want to change the background. So far .8 allows you to swap backgrounds on Nova Reality AM.
Description
kaevee, strawberry-blonde hair, front sections pulled back, loose strands framing face,jedi robe, full-length hooded duster coat, long tan suede open robe, floor-length flowing cloak,layered tunic, wrap-around beige vest, cream-colored undershirt
Comments (6)
Awesome model! Any chance you can make a Illustrious model of her and your Queen Talia model?
Thanks! She came out perfect!
You don't actually need to go through the trouble of making synthetic data to work with polygonal video game screenshots: Just tag your dataset "3D" and you'll seal away any inherent style behind that tag. I did it for my Mission and Juhani LoRAs (plus a bunch more for other games/CG animation) and it works just fine bar some very minor issue (low poly induced straight lines shine through in a few places and Mission's Lekku are very static, and the first could probobly be fixed by adding the "low poly" tag to certain closeups).
As for the background: Have you tried changing the player character's model to her in a save editor? I think it works and it should allow capturing her against any background in the game and in a greatly improved variety of poses if you do. It's also something I did for the above two and I think it works for NPC models (at least humanoids). I'm not familiar with the intricacies/stability of doing it with non-party member models though. (It should definitely work for the party member models though if you want to improve those that way. Only caveat there is meeting the actual character will often have them be headless/invisible since the game is programmed to look for a fallback head if the player character is using a particular head already but none of the unique headed NPCs have one assigned)
I haven't played around with the save editor. I mainly felt like sprucing up Kaevee cause the artwork on her is definitely a little lacking (she is after all an unfinished remnant that is only available cause modders made the unfinished work available). I did train one version based on the original screen caps of Kaevee as well. Talia and especially Mira came out all right from just game screen caps. One of things that turned out to be amusing with the Mira model is that brute forcing with enough images seemed to produce a better result, my last training set being nearly 100 images which is kind of excessive for Anima..
@carneil1000 I find capturing screenshots of the MC at enough varied angles, poses, and BGs for a third of that dataset size more work than I'd desire (It's easy, but its hardly fun.). Can't imagine 100 for a single character, single outfit, LoRA being pleasant. I got Mission and Juhani working with less than 30 and only needed that for including multiple outfits (though with Anima, I needed to lower dim/alpha to 8 to keep the style out even with a tag of "3D". I've heard Anima just wants those lower in general for non-styles.)
@NanashiAnon I ended up using your advice on my second run of an Atton lora and setting the network/alpha dim to 8 and lower resolution seems to work, at least that makes sense when when dealing with an old game's graphics. So thanks for sharing that, also makes the training time considerably shorter on colab.