Hello everyone!
All models are exclusive to Civitai ! Anyone who publishes my models without my consent will be reported!
Trigger Word : Anime style
I highly recommend loading my lora models in fp16 if possible
Since Flux's anime style is a bit weak, I decided to tackle it for my first training of Flux LoRa model.
v3_Final : This version uses the same dataset as v3.1 but this time at 100 epoch against 16 before.
The poses are more dynamic and the characters more expressive
V3.1 : A new dataset of 330 images. Better selection so that characters look older if requested.
The overall quality is also much better
V2_Final : same as v2_15k, but the style is much more pronounced
v2_15k : This version has a dataset of 338 images against 112 for version 1.
This is not the final version of this v2, but I find this 15k steps version very interesting and rather stable.
This version was trained with the optimizer "prodigy" in rank 32
The style is there, but the final version will be very different!
V1 : I used for this 112 high quality images generated with MidJourney v6 in the --personalize style.
Be indulgent, this is a first try. If you zoom in a bit on the images the lines can appear a bit pixelated. But the style is pretty cool!
For a pronounced anime style, Sweet spot between 0.90 and 1.30
I recommend you to play with the strength to vary the pleasures for example at 0.5 or 0.6 by asking for a realistic image.
The model was created with Ostris script which you can find here :
https://github.com/ostris/ai-toolkit?tab=readme-ov-file#flux1-training
Have fun! 😊
Description
FAQ
Comments (13)
Damn you and this LoRA! Witchcraft!
Thank you very much! This is just an early steps checkpoint to give a taste of what's to come 😊
@Stan_Katayama Dear god no!
Strange thing is, i usually avoid anything Anime related, not because i don't like it, i just that if you see it everyday everywhere it's getting boring. But, that thing here creates a lot of random and clean looking stuff, which is pretty nice... :)
@TijuanaSlumlord I try to make models that are creative!
I like to be surprised by results that I don't necessarily expect!
Thanks for your support 🙏
@TijuanaSlumlord Don't hesitate to add more, I love what you do! :)
@Stan_Katayama Sure thing. But, have to keep it down a bit. I like to switch from one thing to something else or it's starting to stagnate. Also wanna finish my own LoRA's so i can create more shenanigans. Flux needs hard love to make it produce stupid stuff.
But, since i like to mix models to see what happens, possible stealthy cross-post might happen :)
@TijuanaSlumlord I'm testing Vantablack right now!
I'm trying combos with my different models.
Yes I know what training can be 😅
I train on civitai but also locally with an RTX 3090. My last local project took me 10 days 600 images, almost 60k steps.
The model is not yet public. I have a lot of tests to do.
Good luck
Hello. What parameters did you use to train it? I still don't understand the model. And I already lost like 4000 Buzz training (If not more)
How should I adjust the Dim and Alpha and everything else?
By the way, very good model
Hi!
I set dim to 24 and alpha to 12. You also need a certain number of images, between 10 and 25 for a person. For a style you need more. 50, 100, ... You also need to put a lot of epoch if you want your model to be well trained.
An example: your model can be very good at 16 epoch, a little less at 21 but become good again at 26 or 31.
But all this is difficult.
You should perhaps try to see on YouTube if there are any tutorials on training lora for Flux on Kohya and try to see the parameters they put.
I am lucky to be able to train in the occale too.
So I test things, experiment.
Everything is not yet defined
@Stan_Katayama Oh I understand. Thank you very much.
@Stan_Katayama Hi! How much gpu vram is needed for local training(flux.1 on Koyha_ss)? I have a gpu with 12 gb vram that I usually use for local training with pony v6.
(Great LoRA!)
@shiroppo I know that it is possible to do trainings with 12GB of vram with Kohya by making concessions on the quality and duration of the training time.
But I do not know the exact parameters
@Stan_Katayama By Kohya, do you means civitai own trainer? Also can you better explain to what Dim and alpha means when training with Flux? Say I what to train a woman for portraits and the like, and have about say, 10, 30 or 50+ images. How does Dim and alpha come into play. Does help with consistency?