Hi :)

I'm currently only taking commissions on ko-fi but if I have time I can consider doing for Julie since I like her too.

Were you referring to the original Julie or the one from New Vestroia?

@AIShitPoster I use the hollowstrawberry colab for most of my LoRAs and recently even use the related dataset editor by hollowstrawberry to tag my images with captions. The LoRAs shouldn't take so many hours, maybe your settings are too demanding? After I've done captioning images, removing unwanted stuff, enhancing the images cropping them (that takes the longest time for me) them it's a matter of uploading them and picking the best settings (takes 15 mins) then I will just leave the lora to train and that part only takes 30 to 50 mins max.

I'm gonna give you the instructions I gave someone a few months back, on how I do my LoRAs. I've changed a few minor things but I still follow the same process almost entirely. I'll put it below

@AIShitPoster this is what I told the person, "Before I explain specifics I want to say I largely followed the guide by HollowStrawberry (https://civitai.com/models/22530) and used their Google colab for training (https://colab.research.google.com/github/hollowstrawberry/kohya-colab/blob/main/Lora_Trainer.ipynb). I did certain things differently so that's what I'll largely explain even though it may not be the best.

I personally find having an optimal dataset as one of the more important processes, so I honestly spend most of my time on it. I'll explain this step first.

1) Creating Data Set

1.1) Sourcing images

I usually get no less than 70 pictures and occasionally use about 200. I number depends on the quality, variety, character and more so there's no rule I don't think.

I prioritise pictures that clearly show the character but make sure that many of the pictures are varied. So I make sure to have them in different clothing, different facial expressions, different angles, different backgrounds.

My personal experience showed me that including nsfw pictures helps the ai know exactly that their clothes in the pictures can be removed/changed and not "stuck" permanently to the character. Having nude pictures is essentially the same as having the character in the "nude clothing style" but I think it helps because there's a lot of source images to compare the nudes with so it better understands the shape of the character. I could be wrong about this though, I'm no expert 🙃

1.2) Editing images

Training LoRAs is great because you don't have to focus so much on cropping images like you would for a textual inversion because LoRAs use something called bucketing. With bucketing you can use any of your pictures as is.

However because I don't know how bucketing works exactly and I know my images are cropped to a square during training. I manually crop pictures to a perfect square provided the character fits reasonably well in it.

Additionally, the longest part is manually going through each picture and cropping out thing I think the ai may get confused with. I remove people too close or touching the character in case it gets their arms mixed up for example. I remove watermarks, text and signatures using ai removal tools. (I use my phone's built in photo editor since I can remove unwanted objects. It's tedious though)

Rename your first picture as "project (1)" where "project" is whatever you want to call your lora. Then the next pictures will be "project (2)", "project (3)" etc. I use "file tools" an app for android to do it, but windows has the feature built in I think, just search it.

Convert png pictures to jpg because I think some colab don't work with it or something.

1.3) Captioning

Open stable diffusion and load a tagging extension, I use this ( https://github.com/toriato/stable-diffusion-webui-wd14-tagger.git ).

Before captioning all pictures, I load and interrogate one picture with the character with a threshold of 0.25 using wd-v1-4-swinv2-tagger-v2. I then find all tags that are going to be essential to the character in all scenarios. Eg: Gwen has orange hair and green eyes would always be used when generating Gwen so I add those to the list of tags associated with the character.

Then add the list of tags always associated with the character and put them in the box for "tags to be excluded" essentially, we don't want any of our captions to contain any of those tags. Then add a trigger prompt in a box to append tags to your captions, eg: Gwendolyn_Tennyson. That way, all those now missing tags in your captions will be entirely represented by that one trigger tag.

Now you can batch process all captions for your images and get your txt files.

Save your txt files and place them in the folder where you keep your images and they will correspond to the pictures they're named after. This folder is your dataset.

2) Actual Training

Going to say this quickly because this is fairly standard and not too different from HollowStrawberry. Its pretty much the default settings in HollowStrawberry, but here's what I change.

> number of pictures multiplied by repeats should be between 200 and 400. So I usually just use the number 300 as my goal to aim for. For example if I have 75 pictures, then I'd take 300 divided by 75 to give me the needed repeats. In that example it's roughly 4 repeats. HollowStrawberry explains this.

> max_train_epochs = 20, though in many instances an epoch of 10 was actually fine too.

> I usually change my unet_lr to 1e-4 because I have a lot of pictures.

> make sure to turn "flip_aug" off if your character has asymmetrical stuff you want to keep asymmetrical.

> I change network_dim and network_alpha on occasion but usually leave them as 16 and 8

"

@Khanarch lol wait please explain why you don't want me to make one of Julie? 😂You are the first person I've seen to beg me to not make a lora