Type | |
Stats | 101 152 |
Reviews | (26) |
Published | Sep 23, 2024 |
Base Model | |
Training | Steps: 468 Epochs: 10 |
Usage Tips | Clip Skip: 1 Strength: 1 |
Trigger Words | Fallout_Head, claymation, clay (medium), portrait closed mouth, expressionless |
Hash | AutoV2 704D9CB302 |
Trained on 32 talking heads from Fallout and Fallout 2, plus 6 ending stills that display the heads (Overseer, Marcus, Arroyo Elder, Myron, Darkwater, and Gizmo). In the original two Fallout games important characters had busts custom made for them out of clay, which was then 3D scanned and rigged to create moving, prerendered, 3D portraits of them. While an iconic part of the duology and something the otherwise complete new content mods with lack, nobody seems to have done a LoRA for this yet. Wonder how it comes out… Post your results so I know.
Tags: Fallout_Head, claymation, clay (medium), portrait
Possible negatives (or positives if that's what you're going for): Old, old woman, old man, mutant, monster, bad teeth, horror (theme) (last one was suggested by autotagger for The Master, some mutants, and Myron, and I couldn't disagree)
Interestingly named tag with 200 uses I found when making this (and used multiple times): corrugated galvanized iron sheet
The training data primarily consisted of neutral expressions and the model seems a bit deficient at facial expressions, so you may want to use “closed mouth, expressionless” and/or avoid strong expressions. The LoRA doesn't even try to generate non-portrait (seems to have no effect on gen that shows more than head and shoulders). Character LoRAs may need to be put to 0.80 strength and (if applicable) source_anime in negatives to work. It's got enough detail on the face that tagging for ethnicity ("dark-skinned female, Indian", "asian, japanese_(nationality)") might also help (I did tag Tandi and her dad for ethnicity). Try saving a copy of your generation as a low quality jpeg if you want to really capture the mid-90s prerendered graphics asthetic
Tell me if you’re interested in a Flux version. I've got some extra Buzz and Flux apparently works better with a limited dataset like this, but I don't have so much Buzz my personal use out of it alone is sufficient cause and will only do it if people will actually generate (and share gens) made with it.