600 steps is all you need (NSFW inside)

Imagine you have a folder of 600 nice blowjob pictures that fill you with a sense of pride because every time you look at them, you think, "Damn, I have great taste in blowjobs." Naturally, you want to create a model that can generate an infinite amount of such amazing bjs.

And because you listen to the shit I say, your dataset prep is just writing these four lines in your Kohya TOML config:

[[datasets.subsets]]
image_dir ='/omnomnom/'
class_tokens = 'blowjob'
num_repeats = 1

Instead of generating 500 useless words per image via Joy Cap, you want your result to be good nonetheless!

Let’s say you’ve got a 4090, and that’s why you’re going for this resolution and batch size - because it fits snugly into your VRAM:

[[datasets]]
resolution = [512, 512]
batch_size = 4
enable_bucket = true

How many steps would you need to train your LoRA for it to converge?

Let’s also assume a learning rate of 1e-4 and dim/alpha of 64/64. Before you start asking dumb questions in the comments I won’t read anyway, here’s the summary: 600 images, one concept, batch size 4, LR 1e-4, 64/64.

No, you don’t need to do some advanced math or shit. Just a ballpark estimate. So you think, "Hmm, at around 2k steps, the concept probably starts to develop into something sensical, and maybe around 5-6k steps, it’s getting somewhat usable. By 10k, it’s really getting there… So, probably a day on an A40 and 12-14 hours on a 4090."

Well, how about 150 steps/20 minutes is enough to get into the "somewhat usable" territory?

You can download the 150-step model here: this_150_step_model_is_better_than_your_4k_step_model.safetensors