The aim of this LoRA is not to be a perfect standalone realism LoRA. For that a dedicated finetune would usually be better. Instead, it's a potentially useful tool to enhance the realism and fur detailing of your favourite checkpoint.
Krea 2:
I was 50/50 on whether I should make this one and post it, but I've been happy with some of the results. Does realism the best out of all the versions which is somewhat expected, but it's not perfect.
Anima:
First attempt at an Anima version. Anima obviously isn't meant to be used for realistic images so don't expect pure realism just yet.
Illustrious:
I wish I can say V2 is a flat upgrade over V1, but they both have there perks. Here's a summary. Feel free to test and choose what works for you. For both I still highly recommend using a face detailer.
v1.0: Intentionally overtrained on a small dataset (mostly wolves and tigers) to maximize realism, but at the cost of some usability. Basically high risk, high reward. Use at 0.6 weight. Higher weights can be done but at your own risk.
v2.0: Doubled the dataset size to include more species, also trained on more reasonable settings. This is the safer option but isn't as realistic as v1.0 as a result. Use-able up to 1.0 strength and beyond.
The reason the lora has "Delken" in the name is because that's the AI director who's images predominantly make up the dataset. I would love to have a more varied dataset but I haven't found anyone who consistently matches the quality Delken in pure realism IMO. As of v2.0, I've started adding my own Chatgpt generated images, but I'm still more than open to suggestions.
Delken if you're reading this, thanks for the awesome pics and sorry that I stole them. Feel free to DM me if you want the name taken off the lora or for anything else.
Description
Trained on 269 images:
{
"engine": "kohya",
"unetLR": 1,
"clipSkip": 2,
"loraType": "lora",
"keepTokens": 1,
"networkDim": 16,
"numRepeats": 5,
"resolution": 1024,
"lrScheduler": "cosine",
"minSnrGamma": 5,
"noiseOffset": 0.1,
"targetSteps": 6725,
"enableBucket": true,
"networkAlpha": 8,
"optimizerType": "Prodigy",
"textEncoderLR": 1,
"maxTrainEpochs": 20,
"shuffleCaption": true,
"trainBatchSize": 4,
"flipAugmentation": true,
"lrSchedulerNumCycles": 3
}
FAQ
Comments (4)
It's so damn cool to see your LoRA evolving! 🤍 Keep it up, bud! You're bring the pure awesomeness, man! Shout out to Delken! 🐺
Some quick info on the Anima version. I forgot to add the "@" to the D3lken tag when training, don't think it affects anything but just keep in mind the trigger is D3lken still. I must've changed my mind about which epoch to use like 5 times hence the delay. Some were better at realism and some were better at anatomy. I ended up going with this one which I feel was a good middle ground. Unfortunately that means it isn't a pure realism LoRA just yet, more like a realistic illustration style. It also heavily changes depending on what checkpoint you use, base Anima looks a lot darker and more raw for example. There are 2 main issues with it atm. First is hands, feet and peens tend to look a little wonky. Secondly, the annoying black border issue has cropped back, luckily its super rare it seems but if you do see it then I apologize.
definitely hard to call it fully realistic.
Still, the fur and scale rendering feels quite realistic instead of being overly simplified or cartoony :)
@BlurHorizon 100% agreed. As much as I'd love it to be pure realism. Just it not being cartoony is the best I can do for now, which hopefully is a breath of fresh air for those using Anima. Thanks for the awesome pics as always by the way!