This Lora does not depict any specific person.
Made using Civitai's Lora training service.
It's my first trial for making Lora. Any constructive advice would be greatly appreciated.
Description
FAQ
Comments (7)
Can you share what training parameters/configurations you used? i keep botching mine and they end up looking too western even though all my training data are asian
Hey there. I just followed the default for the advanced training settings in Civitais' generator.
Model Type: Character
Images: 21
Base model: Semi realistic
Epochs: 10
Num Repeats: 10
Train Batch Size: 6
Steps: 350
Resolution: 512
Lora type: Lora
Enable Bucket: Yes
Shuffle Caption: No
Keep Tokens: 0
Clip Skip: 1
Flip Augmentation: No
Unet LR: 0.0005
Text Encoder LR: 0.00005
LR Scheduler: cosine_with_restarts
LR Scheduler Cycles: 3
Min SNR Gamma: 5
Network Dim: 32
Network Alpha: 16
Optimizer: AdamW8bit
Optimizer Args: weight_decay=0.1
On a side note, have you tried generating images with your Lora using a more asian oriented checkpoints (chilloutmix, realasian, majicmix)?
Thanks for sharing!
Do you also mind sharing what captions you used and whether you used any class tokens/trigger words?
I did try with chillmix and photogasm on automatic1111, but both resulted in a really cartoony output.
I'm honestly not sure what i'm doing wrong here. I was even following similar parameters to yours and manual captioning (not even using blip) eg a photo of kncr1 woman, rest of the description. I'm also using photos from public datasets, so these should be clean and of good quality.
@mornfatstr426
I used Blip for captioning, included the trigger word 'MCInf' in every caption. One example of the caption it came out with was:
"MCInf a woman posing for a picture in a white tank top and black jeans with her hands on her hips and a smile on her face"
I generate my images with majicmix. But i tried with chilloutmix as well, it turned out alright. I am also using Automatic1111
Wow i really like your LORA, what datasets did you use?
I saw the post by another person and saw that you used civitai's training. But same as them I can't seem to get it right, my output always look uncanny.
Do you think that you can share a training guide (other than using civitai's I saw you used automatic1111 and majic model)? I'll most definitely look at it.
Hi there.
I got my dataset from google image search. Influencers usually post high quality pictures, so it made my job easier. I include as many variety of poses, colours and clothing as possible; and of the different orientations (portrait, landscape and square). I chose images with normal lighting, avoiding images that are too bright or have filters applied.
About your output, are you referring to the sample images that Civitai gave when choosing from which Epoch to save? If yes, those images look bad because it doesn't use any negative prompts. I chose the least terrible looking from the bunch, downloaded it, and tried generating images in Automatic1111 with my usual prompts. It came out much better.
I don't dare to make a guide yet, it's probably a beginner's luck. This one is just the second Lora I've made, and the first one was a disaster.
@amru Ahh I get it now.
Your point about sample images being bad seems reasonable enough.
Honestly, your second attempt is already way better than what I could produce so far. I've been attempting to both dreambooth and Lora training for several weeks now, but no luck. Even followed youtube tutorials, tried different captioning styles (using tags separated by commas, being simple rather than descriptive, vice versa).
Details
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.