Before we start: Check out all versions of these LoRAs, each version is a different face.
Who is the girl??? She is No one. You might think she looks like someone you know but she is not.
There are a few things you should know before using these LoRAs.
1: DO NOT USE THEM FOR NSFW OR COMMERCIAL USE.
2: These LoRA files are small in size but if you use them properly they will work really great.
3: Best way to use these LoRAs is to mix them and get interesting results, like using multiple LoRAs in additional networks and giving them different strengths. Eg. No-one1:0.25, No-one2: 0.35 etc.
4: These work great with Hires.Fix if you are trying to get a medium or full body shot, otherwise it will come with disfigured eyes, nose etc.
5: While using close up or extreme close up, lower the upscale in Hires.Fix or it will distort face.
6: These are somewhat flexible LoRAs which work great between CFG 2-10, but you will have to play with sampling steps, sometimes lower steps give good results, sometimes higher steps give better results. In general if CFG is low then keep Steps around 30-90. If CFG is high then keep steps around 10-60.
7: Sometimes fingers and hands turn out weird, better to avoid them in Positive prompt and be sure to add them in negative prompt.
8: Model used for training LoRAs is Protogen 3.4, and for generating images I used MoistMix V1. It should work with most of SD 1.5 models.
I would really like to see what you make with these models.
Please feel free to Post your creations and suggestions.
I hope you enjoy using these as much as i do.
Description
FAQ
Comments (10)
very nice work,
have any details you give regarding dataset images #, regulation images? and kohya settings?
Sure if you want i can provide you kohya settings file (.json). The dataset i usually have is 12-25 images. No regulation image. Keep Network Rank and Network alpha at low (to make small size lora file). I keep 8 bit adam unchecked. Though it does speed up the file making but someone had suggested that it also makes weird far focused images. I tried to keep balance between small size model and its flexibility. But play around with the settings and find what best suits your needs. If you want any help, just let me know.
@Bolly_Diffusion i appreciate the insight, i would like to take a look at your json. im down to chat on discord: S6yx#3618
@Bolly_Diffusion json file would be amazing.. I think you can upload it here as a config file in your model. Then we can grab it to learn. Much appreciated.
Great work on these
@I_dont_want_karma_ Sure i will post .json file but i use old version of kohya so you might have to modify accordingly.
while I don't need lora faces myself, i am really interested in a way of getting my Loras smaller...
any chance you could point me in the right direction?
faces look nice, btw.
Sure buddy, If your aim is just to make LoRA smaller in size keep Network Rank(Dimension) and Network Alpha low. For 4.6mb they both were kept at 4, for 9 mb files they were kept at 8.
@Bolly_Diffusion thank you. I'll test how it compares for my Loras.
she have crossed eyes ? lmfao
Which one? there are 7 of them, scroll down and check versions. Also they are digitally created anything that you dont like can be changed with negative prompt or inpainting.
Details
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.