Hinata / The Byakugan Princess
Trained on 147 images. This is my fifth Lora and I feel like this might be the best one i've train so far. Use weight between 0.65 and 0.9.
Look i do not take request just to be clear. and i'm not gonna make every single naruto characters. I'm just here to make some Loras of the characters that i like in the series.This might seem selfish but i don't have the luxury of time and electricity.Hope you understand.
If you appreciate my work please consider leaving a review.
i used training reso 768x768 for this one i feel like it is more detailed. But due to the limitation of my hardware have to wait for several hours. But worth it.
Here is the basic prompt for the original outfit:masterpiece, absurdres,hinata\(boruto\), 1girl,mature female, solo, purple hoodie,layered sleeves, brown pants, outdoors, smile, looking at viewer, cowboy shot
Here is a bonus for her outfit with her black dress and apron:masterpiece, absurdres,hinata\(boruto\), 1girl, solo,blush, black shirt, apron, looking at viewer,kitchen, smile`
Description
FAQ
Comments (21)
I will test it soon, your LoRAs are good.
Is it very difficult to make these LoRAs? I'm saving some change to buy a better video card to start making mine.
@MySmartArts Thanks for the compliment. Yes it's a time consuming process. Especially the data collection and tags pruning or image labeling process. if you are going to buy a GPU make sure you buy the one that have a lot of VRAM if you have the laverage. Trust me it'll make your life so much easier in Lora training.
@Rindo192 Is the RTX 3060 with 12Gb VRAM enough?
@MySmartArts Yes
I'm a huge fan of your loras, you do a really great job at isolating the subject with your tags so that the training algorithm gets a good idea of what the character should look like. I notice in your examples you never have to call up the look of their eyes or hair or anything like that. Truly a perfect example of how a lora should be done!
Just a suggestion since you tend to do Naruto characters: you should consider training all of them into one lora. I'm not saying you shouldn't have each lora available separately still, but if you run the training algorithm on all of your data sets at once (keeping each character still in her own folder/s), you can create a lora that is trained on all of them at once. This becomes interesting when using Latent Coupling. Latent Coupling does not tend to like when more than one Lora is loaded, so creating fanarts with multiple characters is hard. With a single lora trained on multiple characters, though, you can get each character to show up in a latent coupling gen no problem.
I know it would likely take a lot longer to train a big lora that has all of your Naruto characters in it, and you'd have to update this "mega lora" any time you create a lora for a new character if you wanted to keep it on parity with the individual loras... but the advantages are huge, in my opinion. I hope you consider it.
@Knots Well thanks for the compliments. I do take out most of the tags which are implicit to the character thus those features are baked into the Lora but at a cost of flexibility. My intention is to make lora with little prompting as possible and make sure that the trained character actually represent the original anime.
As for your suggestion i have thought about it a while ago to make all the akatsuki members in one single Lora but i'm still learning about how to make a loras. And i feel too foreign about training multiple characters. May be in the future when i have a beefy PC i might consider it. I never knew that you could gen multiple characters with latent coupling if the lora have multiple characters in it. Truly fascinating.
Anyway thanks for the info.
@Rindo192 I appreciate the consideration. To explain further how you might combine your loras, here is the case of combining Hinata, Sakura, and Hinabi loras using your existing datasets.
Rename "NSFW" folders for all 3 characters to "8_Hinata_NSFW, 8_Sakura_NSFW, 8_Hinabi_NSFW"; this is to avoid combining their NSFW datasets into one "concept" since the folders are called NSFW across all 3 loras originally.
Then, put all of the data for all 3 characters into the same /image/ folder. You'll have 9 folders this way. In this example, I also put the character names in the names for the other folders for the sake of staying organized, but it would look like this:
8_Hinata_NSFW
8_Sakura_NFSW
8_Hinabi_NSFW
15_Hinata_hoodie
15_Hinabi
15_Sakura
20_Hinabi_Hana
20_Hinata_blackshirt
20_Sakura_Uchiha_crest
Then, you should be able to just run the training with the same settings you used originally. It will take a long time because you're training 3 times as much data as normal. The results should be indistinguishable, though. In my tests, I trained all 5 Madoka Magica character loras together into one lora and the results were very similar to their individual loras when I prompted for just 1 of each of the characters.
I hope that helps you understand what I mean. I understand it's asking a lot of your hardware to do a big fat lora like this, but if you decide to do it now you know how, or at least so far as I understand how.
@Knots That's a well thoughtout advice. I'm flattered that you know how the i organized the datasets. i feel somewhat exposed ngl XD.
It took 2h to train the Hinata Lora with 768x768 with the batch size 2. The other loras cost the same amount of time more or less. if i were to train 3 of them as you suggested it might take like 6h.
But the thing is that in my country i don't get like 24h electricity like 4h on and 4h off and most of the time that is not always true it can be random.
I'll try and test the process as you described when i have my free time, if it's possible. And Thank You for the advice it help me understand better how multiple characters Loras work.
@Rindo192 lol every lora comes with metadata on how it's trained at the top of the .safetensors file so it's not hard to figure out. You can open one up in a text editor and see for yourself, and Automatic1111 also has a way to read the metadata by clicking the "i" button on a lora card in the additional networks tab :P
Wow, that sucks. I'm sorry to hear that about the power issues. I would be willing to run the training for you and any other datasets you wanted to run if you like, though I understand if you'd prefer to do it on your own. If you decide to take me up on the offer, you can find me on discord at Knots#7202 or you can email me [email protected]
Either way, keep up the awesome work. I wish every lora on Civit was as consistent as yours are!
mei terumi shippuden please
OMG!The best hinata lora!shippuden version,please.
Hey ! huge fan of your work ! Is there a way i could DM you by any chances ? Thank you for your answer and keep up the awesome work =D
I really love it, waiting for more akatsuki!!
Really good, thank you!
Hinata from Last Naruto?
hello , is theres any tutorial on how install stable diffusion and make fanarts like urs plz ?
Top of the top!!! Would you update it with some of her jutsu? ^__^
@Romazeo I believe that in Boruto she doesn't do any combat activities. But if you are referring to her abilities like gentle fist, or twin lion fists. it could be done but only supported on the Shippuden version. When i last checked the dataset is kinda low for the jutsu So i left it out on purpose. May be i might try adding it one day when there is enough dataset.
great lora!!!
pls do Tsunade also, love ur work!
很棒！感谢分享
Details
Files
Hinata[boruto]V2.safetensors
Mirrors
HinataborutoV2.safetensors
hinata(boruto).safetensors
Hinata[boruto]V2.safetensors
Hinata[boruto]V2.safetensors
Hinata[boruto]V2.safetensors
Hinata[boruto]V2.safetensors
Hinata[boruto]V2.safetensors
775_Hinata[boruto]V2.safetensors
Hinata[boruto]V2.safetensors
Hinata[boruto]V2.safetensors
Hinata[boruto]V2.safetensors