Type | |
Stats | 142 4,736 |
Reviews | (34) |
Published | Dec 25, 2024 |
Base Model | |
Training | Epochs: 24 |
Usage Tips | Strength: 1 |
Trigger Words | gigachad \(meme\) the chad smile the serious gigachad sitting gigachad lean |
Hash | AutoV2 3E4C2EFAB7 |
Added a version without the IterComp weights, for use with already itercomped checkpoints like Cat Tower or if you just don't like it.
With the release of Noob vpred 1.0, and taking bits of advice on training settings, and refining the dataset over a couple generations, I decided to retrain my new loras and split this version of Chadmix from previous ones. It's an aesthetic balance set built around a Gigachad meme dataset, which I happened to notice was enhancing character details. And this bake is a juicy one.
Noob Vpred 1.0 has a significant learning curve unto itself and this doesn't do anything to change that. It doesn't provide a strong default style for the model to fill in details more consistently on short easy prompts. It reins in colorspace blowouts, biases gens towards more detailed characters and backgrounds, and provides Gigachad meme knowledge. I've gotten better results pushing CFG as low as possible on a CFG++ sampler, as of writing I'm rolling with 1.07 CFG on ComfyUI's scale, the lowest I've ever gone.
The booru tags for gigachad \(meme\) etc apply, and I added "the chad smile", "the serious gigachad", and "sitting gigachad lean" as additional tags to prompt their respective poses/expressions.
The recipe starts with a Noob + Itercomp merge, then I trained on top of that for 24 or so epochs, then extracted a 96dim linear 24 convolution locon using base Noob as the extraction difference. So it can potentially conflict with Itercomp mixes.
I also recommend using this with my updated Chiaroscuro lora as a light/shadow/contrast control knob.