V2 is out! The update improves the model's consistency and stability. It is more authentic to the anime art style. and it can draw her hair color correctly most of the time. The sharpness issue from v1 has also been fixed. This model has exited the prototype stage.
This is the flux1 dev LoRA of Dr. Akagi, the beautiful blonde hair scientist from EVA.
As shown in the showcase images, you can draw her in different outfits and settings.
And flux still doesn't know how to draw an EVA unit...
Description
FAQ
Comments (14)
BTW - obviously flux1 does not know what an EVA is, but knows Gundam quite well.
Did you train this here on Civitai? Can you please share what a sample caption file looks like? š
I trained on Civitai, yes, as it is too slow to train it on my own PC. I reused the exact same caption files as SD1.5 though, using WD14, removing all the intrinsic tags such as "short hair" and "blonde hair". However, I almost suspect flux1 dev LoRAs will do just fine without tagging, and tagging probably didn't help with the training process.
At any rate here is a sample caption for one of the images:
"ritsuko akagi,akagi ritsuko, 1girl, solo, breasts, bangs, large breasts, simple background, white background, holding, bare shoulders, jewelry, upper body, earrings, sleeveless, shiny, off shoulder, lips, makeup, gem, smoke, cigarette, labcoat, smoking, red lips, holding cigarette"
@phageoussurgery439Ā THANK YOU SO MUCH!!!!!
One more question: what was the resolution you used in the dataset? 512 x 512? @phageoussurgery439
@wktraĀ In training I used 1024x1024. The dataset itself are of various resolutions though. I tried to use higher resolution images whenever possible, but many EVA anime screenshots were 620x480 or lower - they still work.
@phageoussurgery439Ā that's awesome. I'm INSPIRED by you!
How much did it cost to train the lora on civitai? Did you do several attempts?
@wktraĀ it's 2-2.5K buzzes, depending on the number of epochs. So it's like $2.50, which isn't bad since I would have to toast my own GPU for a few hours otherwise. Believe it or not - half of my trainings were to get a working Asuka š. She was difficult. All other trainings I was able to get satisfactory results in one pass.
Again, I think tagging isn't too important for Flux anymore. And flux seems to be able to tell the character apart from their outfits, so no need to worry about over fitting on the outfits. In someway, flux can be easier to train than SD1.5.
Looking forward to your models!
@phageoussurgery439Ā I guess $2.50 isn't so bad. š¤ Was it just getting the tags right that made you have to go through all of those failed trainings? š²
Any tips to save me some cash? š
@wktraĀ Asuka was just difficult for unknown reason. I was trying hard to avoid flux's built-in anime style as well.
But otherwise, my tips:
- collect as many high def images as possible, ideally higher than 1MP (total pixel higher than 1024x1024), which will help when you train at that resolution.
- again, I don't think accurate tagging matters much. Just use WD14 tagger and remove the intrinsic tags (I typically run WD14 on my own computer twice: the first time to discover the high frequency intrinsic tags, then add them to the tags blacklist and rerun the tagging).
- I think you want total steps to be more than 1K, just in case it takes longer for the LoRA to be fully baked.
- I set network dimension to 32, alpha to 16. Clip to 1 for realistic and 2 for anime. Everything else is default.
Good luck!
@phageoussurgery439Ā thank you so much šššš I've made enough tips at the nail salon this week to do a few batches of training! š¤š
@wktraĀ Have fun! BTW I managed to tackle Asuka - posted a newer verrsion that is much easier to work with. Check it out if you like!
@phageoussurgery439Ā doing it now! š Thank you!
Details
Files
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.