Honovy is an artist with a strong and distinct style. Many finetunes will try and fight the style and try clean it up, especially on the illustrious version. IMO Krea 2 captures the style quite well, at the cost of some nsfw capabilities as of writing this.
Description
Training params are below. In hindsight I maybe could've been a little braver with my repeats per image.
281 images.
{
"engine": "kohya",
"unetLR": 1,
"clipSkip": 2,
"loraType": "lora",
"keepTokens": 1,
"networkDim": 16,
"numRepeats": 3,
"resolution": 1024,
"lrScheduler": "cosine",
"minSnrGamma": 5,
"noiseOffset": 0.1,
"targetSteps": 4215,
"enableBucket": true,
"networkAlpha": 8,
"optimizerType": "Prodigy",
"textEncoderLR": 1,
"maxTrainEpochs": 20,
"shuffleCaption": true,
"trainBatchSize": 4,
"flipAugmentation": false,
"lrSchedulerNumCycles": 3
}
FAQ
Comments (5)
I can say that the male sexual organ is forcibly transformed into a dog's. Adding negativity or a strength below 0.8 does not always help me.
The rest needs to be tested, but judging by the samples, it's not bad at all.
Thanks for the feedback. Can't say I've been seeing the same issue, but it's not at all unbelievable that it would happen given the artist uses quite a bit of canine peens. All I can recommend if you haven't tried already is to either add "humanoid genitalia" and "foreskin" to the positive, or to try a checkpoint that is slightly less furry-focused like WAI.
@Boerboer It seems Honovy likes to make the images very noisy and blurry
@qek Yeah its definitely their signature look. As mentioned in the model description, any lora would struggle to achieve that blurry look as most checkpoints (other than base Noobai and illustrious maybe) are trained to specifically not do that.
Awesome style!