How many steps is too many, how many is too little, and how do I find the right balance?
I'm cooking some LoRAs and I noticed some of the characters have their outfits baked onto them so much that when I give them variations, even the shade of their signature outfits remain, as though they have bleached the clothes.
So I'm wondering how many steps is too many for a LoRA.
For context, the one dataset I'm having issues with is 40 images at 1290 (this is with 6 Repeats and 10 Epochs), but knocking down the Repeats down to five will net me 1030 Steps so I'm wondering if that's too little steps for the cook.
So I'm left to wonder what the right balance is.
Update: Tested out the 1030-step one, no, it doesn't seem like I overcooked it. I think it's just the limited wardrobe the character had just baked itself onto the LoRA.
Perhaps it could be the "BLIP captions" you are using for the training. (?)
These are very important in order for the A.I. to understand what Exactly you want for creating a model.
Are you creating a Lora for a person, a fictional character, clothing (or object) , or an art style ?
I usually use about 36 to 50 images to train a LORA for a model of a PERSON Face and body.
I train my LORA using the original Stable Diffusion 1.5-pruned.ckpt , or the original Stable Diffusion 1.5-pruned.safetensors.
which can be downloaded from here...
It is the big 7.7 GB file.
The reason for using this file is because if you train your models on it, it can be used with any other
Checkpoint being posted here at Civitai for photo-realistic images, except for the Anime-art-style
checkpoint versions which tend to make all the person models look a little Asian or an Anime cartoon art-style.
If you use some other merged Checkpoints for your LORA model training, this may not work properly with
other checkpoints. The USER may need to use the same checkpoint as the one you trained your Lora with.
I usually just need one training and my Lora file usually comes out super great.
Below are the things you just need to write in your LORA training Menu, nothing else is needed.
(At least for me, it works perfect)
LORA type - Standard
Train Batch - 2
Epoch - 1
Caption Extension - .txt
Mixed Precision - fp16
Save Precision - fp16
Seed - 1234
Learning rate - 0.0001
LR Scheduler - Constant
LR Warmup - 0
Optimizer - AdamW8bit
Network Rank - 128
Network Alpha - 128
Enable Buckets - YES
in Advanced Configuration - CLIP SKIP - 2
Max Token Length - 75
Bucket Resolution Steps - 64
their signature outfits remain --- you need to add the color tag in txt document.
1000+ steps should be enough, but you need to increase your LR.
40+ pictures for clothes is enough, If you do not want to increase your LR, you need more setps, why not try 3000+ and 6000+ then save everything you got and make a test ?