What's the best way to begin diagnosing what went wrong with a LORA in terms of being off-model?
What's the best way to begin diagnosing what went wrong with a LORA in terms of being off-model?
GI've followed the tutorial for: Become A Master Of SDXL Training With Kohya SS LoRAs
GMy model is decent but the likeness is off
GBtw I subbed to the $5 Patreon
GOne of the issues besides likeness that I'm seeing is weird distortion on the face. Waves on the face like I'm melting
Fhello ty so much for patreon
Fwhich gpu you have?
Fi strongly suggest SDXL dreambooth instead of LoRA
Fmuch more powerful
Fthe likeliness issue 99% related to training dataset usually or used config
Ahi guys, is there a way to make an ai with the voice of someone else? i'm really curious about this since i saw a lot of videos with ai generated voice of real people and was wondering how do they do it
GGTX 3060 12GB.
GWhich online service is the best for running Dreambooth?
GDoesn't have to be free
Bfree: kaggle, paid: runpod
GThank you 
GIs RTX 4090 24GB good enough for Dreambooth SDXL or should I go with better like RTX 6000 48GB?
BA5000 or 3090 are sufficient for Dreambooth SDXL, works fine with his 24gb settings file
Frunpod
Fbest
Feven rtx 3090 good
GThanks @Furkan Gözükara SECourses
GWhat's the difference between the kohya gui for training Dreambooth and the A1111 Dreambooth extension?
GI am confused as to which tutorial to follow
Fa1111 dreambooth is good for sd 1.5
Fbut only older version working
Ffor sdxl kohya is the one
GHas anyone made a graph of the number of high quality source images for a model and the quality of results?
GI understand that quality is more important than quantity. But as an example let's say... is there any reason to go above a certain number of source images assuming they're all high quality?
GWhat's the diminishing returns like?
Fnot that i know such graph
Fit is mostly based on experience
GGIs there something wrong with the settings in these JSON files?
GI've done training with 24GB_Best and 48GB_Best, and it only creates a single model file, and the model is way worse than when I did the LoRA
G
GThis is for 18 source images btw
GI think I confused this with LoRA. Is the cause of my issue because I put max trained steps in the Regularization Repeats?
GOkay, I rewatched the original LoRA tutorial I got this idea from. When you explained how to calculate how many regularization images you need based on steps and number of source images, I for some reason thought that I needed to put that number in Regularization images Repeats
F5760 repeat in reg images
F
Fit will be 1
MI have the exact same issue. @Furkan Gözükara SECourses are you saying i should not use LoRA to train my face in SDXL? (4060ti16GB)
GI'm re-running my training now after I fixed my issue I had in my settings. If that issue is solved for me I'll let you know.
FI would do dreambooth in your case
FWe have 16gb config
_Is there any difference between sdxl finetuning and kohya's dreambooth? Since both methods use sdxl_train.py?
(except for the captioning method..)
(except for the captioning method..)
Fjust dont use reg images
Fwhen no reg images used dreambooth = fine tuning
_thanks a lot!! now it's little less confusing 