Style based on artist Polki's works, and does very well with canines and equines. Trained on 148 images with ~5h of tedious manual tagging.
Seems to work very well at a strength of 0.8 on base Pony, but you may want to push it up to 1.0 with longer prompts, or on merge checkpoints like AutismMix.
If you like the style/artist, I recommend following them on their (NSFW warning) twitter, pixiv and/or fanbox.
Description
FAQ
Comments (8)
This is very high quality! Just wondering, how to do you go about tagging your images? I'm looking to improve my Loras
Thanks for the praise! I'll try to give a brief rundown of my process, but there's a lot of stuff built up over the past year of fiddling around.
1. Initial tagging is done via the A1111 tagger extension (tag threshold set to 0.4):
https://github.com/picobyte/stable-diffusion-webui-wd14-tagger
I use the Z3D-E621-Convnext interrogator from here to get e621 tags (which is what Pony uses):
https://huggingface.co/toynya/Z3D-E621-Convnext/tree/main
2. I'm pretty well-versed in e621 tagging, so after that it's a manual "tag what you see" to tidy up the autogenerated tags and add new ones, with particular focus on things you want to be able to reproduce when using the LoRA when it's done.
You can use something like this to speed up that process:
https://github.com/starik222/BooruDatasetTagManager
And if you do, it also helps to replace the danbooru.csv in BooruDatasetTagManager/Tags with the e621.csv from here (also a very useful A1111 extension for prompting):
https://github.com/DominikDoom/a1111-sd-webui-tagcomplete/blob/main/tags/e621.csv
3. Also for some training tips, I would enable 'shuffle captions' if you're training a style, and 'flip augmentation' if you have a lot of images which are variations of one picture (like Polki does).
Full training json from civit:
{ "unetLR": 0.00005, "clipSkip": 2, "loraType": "lora", "keepTokens": 0, "networkDim": 32, "numRepeats": 6, "resolution": 1024, "lrScheduler": "cosine", "minSnrGamma": 5, "noiseOffset": 0.03, "targetSteps": 2220, "enableBucket": true, "networkAlpha": 32, "optimizerType": "Prodigy", "textEncoderLR": 0.00005, "maxTrainEpochs": 10, "shuffleCaption": true, "trainBatchSize": 4, "flipAugmentation": true, "lrSchedulerNumCycles": 3 }@DrA_263 Awesome, thank you so much for your help!
@DrA_263 I'm having a bit of trouble trying to get the Z3D-E621 interrogator to work. It says it has ran out of memory, but I don't know whether it's caused by an improper installation. I just put it in the models/deepdanbooru folder, and created an empty project.json.
@scringler Its an Onnx model, so it needs to go in models/TaggerOnnx/
I'm not sure why you'd be getting an OOM error though, it should just not see the model at all until it's in the correct folder.
@DrA_263 Thanks for writing this guide, it helped me a lot with my own style lora. :)
@ContainmentBreach I saw! I've been using it and it's great, nice job!
Details
Files
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.