>>8680928
this is correct
naiXLVpred102d_final


>>8680934
its the edm2-rebake of h@rypotter lora <lora:harryp_noobv1vpred_edm_v001-000036:0.8>, you bloody basterd

9 baked in chars + scenery.
Examples with meta:

argusfilch.jpg
https://litter.catbox.moe/vi7kpefgcq4yi2wg.jpg
gob.jpg
https://litter.catbox.moe/bi21h4au0n4b9cza.jpg
herm.jpg
https://litter.catbox.moe/5zyg4erp0z5geur5.jpg
hp.jpg
https://litter.catbox.moe/6v243onhh3hd9da7.jpg
severuas.jpg
https://litter.catbox.moe/ny9q1njjzjl7adva.jpg
troll.jpg
https://litter.catbox.moe/m642jtekmhrsf9ll.jpg
rub.jpg
https://litter.catbox.moe/os3pmlg8mm6wv0ep.jpg

scen1.jpg
https://litter.catbox.moe/y33fkrsdzldpdaaw.jpg
scen2.jpg
https://litter.catbox.moe/ki7bxxzptr96lszb.jpg
scenTent.jpg
https://litter.catbox.moe/5xcz3q3yxoir6ror.jpg


dontclick.jpg
https://litter.catbox.moe/noadiy6dhigwiiw0.jpg


my takeaways from the bakes:
-Best to go with 32-36 epochs (for "1_" datasets). Its a sweetspot without underbaking and overbaking.
-If you train on real imagery, tag with "photorealistic,realistic". You can remove those while genning to get a more 2D look. Or use those for more 2.5+D.
-For chars 30 good images are enough.
-Scenery and backgrounds are harder to bake it than the chars and need al least x3 more repeats (if you have only one image of some scenery/location). Actually forgot to rename scenery dir to "5_" on this bake and baked it with 1_ , so all sceneshots except tentacles sucks (had for images for tentacles/veins ;D ).
-SXDL sucks major ass in 2025, no finetune can redeem it.


My LoRA_Easy_Training_Scripts config takes now 6s/it on 67372a fork, but it was 3s/it before :(
Need to check whats going on there...