Hi everyone, as always, all comments and criticisms are welcome, as they help me to continue learning!
If you like my content please don't forget to rate it! ♥
Info: 150-step TI trained on a dataset of 15 images
Based on hopelesssofrantic (Australian Pornstar & model)
Description
FAQ
Comments (6)
150 step?? I use 3000 steps as the bare minimum. Is there something I miss out? Or you're refering to anything else?
Hello, I leave you my article on how to achieve these results with a few steps
https://civitai.com/articles/692
@mv_ia I see. I use a pretty similar settings except the gradient accumulation steps that really slow downs a lot the process, but If you only use 150 steps is not that bad. I'll give it a try. Thanks.
I've been testing long and wide your metholodogy. It certainly can give a good output with an excellent dataset in the range of 150 steps, but I think is too low setting for my taste. I'd rather put something like 250-300 steps for the extra detail/ likeness. At the end of the day, this process is 2x faster than my previous one.
What I've found and probably is the most interesting thing I've realised, that going up in the vectors per token gives you a far better result. I was using "5" but with "10" vectors per token the output is a lot better for faces. It works well with a dataset of 20-40 pictures. (what I recommend for subjects, anyways).
@Punkimaster I was looking at some data based on what you tell me, and I saw that values like these can be used:
2-3 Vectors Per Token for 1-10 Training images
5-6 Vectors Per Token for 10-30 Training images
8-10 Vectors Per Token for 40-60 Training images
10-12 Vectors Per Token for 60-100 Training images
12-16 Vectors Per Token for 100+ Training images
I am going to do new tests using this VPT scale, and seeing how much they influence tests of 100 to 300 steps.
@mv_ia yes I know those data from reddit, I think is outdated. You want to use 8-10 vectors per token for faces and subjects (with a dataset of 20-40 pictures). I think 5 vectors is too low. But check it out by your own. There's no penalty in the training time process so you're not gonna waster more time in the process.
Details
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.