Try size: 640 x 640 or 512 x 768
Weights: 0.8
Clip skip: 1
Training material from bing
Pictures from Calicomix and Animagine XL V3.0
Note: If need ghost, please add "(((ghost))),".
Description
Training from Anything V5.0
FAQ
Comments (17)
👻❤️👻
can you please share the dataset? (there is a button on the right to add files)
I think it would be interesting to see what did you used to train and how it learn it...
Lora from Pixai, It doesn't give me any data...about training checkpoint is anything 5.0. I rarely use kohya_ss, usually realistic Lora.
@ghostpaint Explayn better: didnt you collected images, cleaned them and tagged before do the train?
I mean, for example, here I asked to see the dataset (on the right there is "Training Images - download"
https://civitai.com/models/228559/edob-urusei-yatsura-kurama
@settima_ai I don't have any data...all lora from Pixai...I can’t see any training data in webui either.
@settima_ai 假如你需要討論，我只能在右上方的Chats頻道中使用中文對話，因為我回答英文很吃力，而你可以繼續使用你習慣的語言。比如你想做什麼樣類型的lora，換做是我會怎麼進行…我剛下載過了，那數據太少，只有標記，對你應該沒有幫助。另外自動標記，我是看不到的，我只有下觸發詞，在Pixai只能靠觸發詞控制lora。
@ghostpaint Oh, I dind't notice you were not english!
If you have no data, my request has no sense to exist!
Thanks ayway! : )
nice idea 😄
and by thinking about it you could train for more movie genres like comedy (laughing waifu and cowns or laughing people in the background) romantic movies
If I'm ready...😅
@ghostpaint take your time, those are just some wucik thoughts ^^ it's not a request, so you don't have to do this if you have other models to train
This is great. Please do an SDXL one!
Thanks. I have never trained SDXL Lora, do you have suggestions?
@ghostpaint Its basically the same thing, just select sdxl. Make sure your not using as many low quality images. Its been a while since I did it so Im sure my knowledge is outdated, but I just followed a tutorial I found, cant remember the details.
Thanks for releasing this in SDXL. Would love to continue seeing your other 1.5 ones as SDXL as well
@SnappleRYRY89 For example?
@ghostpaint im happy with everything that hasn't been already converted to SDXL. I preferably like "Girl like parade float" but any and hopefully all would be great
@SnappleRYRY89 At present, unable to convert, must retraining dataset in sdxl. Usually about 70% ~90% similar. I'll try to be more similar...