AI think it's diverse this time
A25 images
A
A40epoch Lora results - It looks promising so far !!! The other models I did wouldn't even grasp the font at 40, epochs
A
A80epoch , 30 steps perfect results. It recognizes the line above now after 20epochs
RLooks really good!
AAlso the more I train the model the less steps the inference requires
A
MSo the dataset changed everything, looks nice!
MI think there should be a "rate my dataset" section here
Ffp16 and bf16 same size and they are half of fp32
Fi never used in kohya yet 
Fyou need to update it
Fsince main branch still not updated
Fi have updater script
FPatreon
Get more from SECourses: FLUX, Tutorials, Guides, Resources, Training, Scripts on Patreon
Ffollow massed compute instructions
Fit also has very fast download models
Fstarting training takes like 3 minutes after initalized 
Flooks way better
ADoctor you and timson were right about diversity
Fyes i really optimized workflow for everything except perhaps ultra big trainings
Fso it is up to your dataset
Fand caption
AI let flux caption
Athis is game changing actually it can be combined too
MWhen generating a dataset for a character with flux for lora training, would developing a certain guideline for poses be beneficial? or random images is better.
I'm thinking let's say: 8 close up images from different angles + 8 full body different angles + 8 waste up different lightning .
I'm thinking let's say: 8 close up images from different angles + 8 full body different angles + 8 waste up different lightning .
MThe reason I'm asking is that if we take this systematic approach, one click image to lora would be possible. because control net can be used to generate very certain poses.
Fyou should have every angle and pose you would like to generate after training
Fthat is the key
Fand also emotion
ADoctor does training on .jpg or .png have any major difference?
Mthen add 8 for emotions, or maybe sprinkle some emotions across the dataset.
Mas if it's that easy!
Fshouldnt have any diff
Ayou can train a stage 1 model and generate pictures of your subject with emotions
Athen train with a new dataset
Aor use inpainting on flux but haven't tested that method
MI am testing Pulid to do so, without having to do multiple stages, it's pretty good
Amake sure no noise is being generated/blur
MIs this considered noisy? (I am still trying to get rid of the plastic skin)
Mthis is a low res, here's the detail:
Fit looks like cgi though
MAgree, still working on that.
FSDXL training now goes to as low as 10200 MB on Kohya
A@Furkan Gözükara SECourses what's the other unique token for sd called other than ohwx ?
Ahere's another shot I made