Attempts to replicate the style of portraits featured in the 2001 video game Arcanum: Of Steamworks and Magick Obscura. I threw all 76 Arcanum portraits on the Arcanum wiki’s portrait page (I think there’s a few more unused portraits) into trainer, applied some tags and saw what happened with it.
Tags: arcanum_portrait, portrait, traditional media, solo, simple background,
Helps: painting (medium)
Aspect Ratio: Square (all originals are square.)
Tagged species: Elf, half-elf, orc, orc, half-orc, ogre, monster, lizard (note: all half species are also tagged their full, so all half-elfs are also tagged elf. Exception is half-ogre is just ogre because there are no ogre portraits or half-ogre tag)
Negative: Source_anime (this is very important)
For best effect: Save a copy of your generation at 10-20% size as a deliberately lossy (15% quality) jpeg.
Recommended Checkpoint: Pony Realism? (Best so far of what I've tested. See gallery for what I've already tried.)
You should go tell Microsoft to let Tim Cain release the Arcanum source. He has it and it compiles, but the only thing stopping him is his contract with Sierra (owned by MS since their acquisition of Activison) that prohibits him from releasing it over concerns it could break the anti-piracy (which was cracked within weeks and the game has been sold without for well over a decade).
As for making this: It was pretty straightforward, except the autotagger was just completely worthless here. It failed to recognize people being present, and when it did it thought men with clear beards were just “1girl”. It took a several epoches to be more recognizable, but I think the big issue is I couldn't negative tag "source_anime" during training.
Description
1
FAQ
Comments (5)
Hey, great job!
I really like how the trainer portrays the game's distinctive style, and also it's pretty cool seeing a creator who has a taste for these lesser-known but great games and their art aesthetics.
Kudos for your work, both with your great stuff here in general and this type of niche idea in particular!
Good to see someone likes it. Please show me what you make with it. You may also like my Fallout Talking Head style LoRA if you like this. I'm still trying to figure out how to dial the settings (including Checkpoint) before I judge this one.
As for the lesser-known stuff: A lot of the stuff I pick to train comes from thinking either "I can't believe this doesn't have a LoRA yet" or "I wonder if this will even work." PDXL has generally surprised me with the weird and low quality stuff I can throw at it and still get a functioning LoRA
@NanashiAnon Yeah, it's a pleasure to provide feedback on a model a like, especially when the creator obviously put in so much effort and puts out such good stuff.
BTW, interesting that you say PDXL has a surprising ability to produce functioning LORAs from low quality source material, I wonder if this connects to the way it seems like its massive database (with detailed tagging) of porn gave it an unprecedented grasp of how human anatomy fits together.
I mean, I'm no expert, and I didn't come up with this idea myself, but my own experience with Pony validates those who I hear saying that it seems like it's a model that actually gets how human anatomy works, in a way that far outstrips its peer models.
@ravemry9 That certainly helps for human figures, but I've gotten actual character specific details pretty well from just including stuff like a handful of heavily compressed cutscene pictures in the training data.
@NanashiAnon wow, that's interesting. And a very useful thing to know; thanks for sharing!
Details
Files
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.