for sd 1.5 i suggest dreambooth extension
for sd 1.5 i suggest dreambooth extension
Fhere the auto installer link and settings
FFFdreambooth extension supporting EMA
FKohya will also hopefully implement it
FI will also research Kohya best settings for SD 1.5 too
MSo the .json file is within the autoinstaller ? it comes with Parameters built in ? since it's in auto ?
I don't see the JSon file if not
I don't see the JSon file if not
Ebetter now clean ram
MI just trained a Lora model on RealisticVision with 32/32 (NR/NA) and Prodigy with these parameters:
decouple=True use_bias_correction=True safeguard_warmup=True
It's amazing to achieve professional results with such low weights on a SD 1.5 model. Here is a sample (original and hires fixed):
decouple=True use_bias_correction=True safeguard_warmup=True
It's amazing to achieve professional results with such low weights on a SD 1.5 model. Here is a sample (original and hires fixed):
Mi do prefer lora Kohya ss and I also started following your video ( produce studio quality images)
I installed the Kohysa ss Gui , now I only need the setup parameters ( if you have any json file ready )
is the one in this link work with kohya ss ?
I installed the Kohysa ss Gui , now I only need the setup parameters ( if you have any json file ready )
is the one in this link work with kohya ss ?
Ffor sd 1.5 i am still in research. i have this but it is not best here : https://www.patreon.com/posts/best-stable-1-5-90307694
Fye
Finteresting i should test it
Fcan you send me config file
MI sent it.
MWhat I leave as a default, I don't set as a parameter, it's unnecessary. I only put in as an args parameter what I modify.
Fhere a golden info
S@Dr. Furkan Gözükara any clue what part of the pipeline is causing my sdxl dreambooth fine tunes to output noise?
SI had it working end to end a few days ago, but I think I broke something. Trying to figure out if the training was bad or if it's an inference problem...
SMy hunch is switching to the fp16 vae. Have you encountered similar noise?
Fovercooked unet
Sok, I'll check the earlier checkpoints to see if it's overcooked
S@Dr. Furkan Gözükara ok getting the same output for all the checkpoints, so definitely a vae issue. How can I identify if it's an issue with the inference configuration vs. the vae I configured for training?
SNot convinced I'm loading the vae or unet with the correct dtype
Fwhich vae do you chose during inference?
SThe one you linked me last week. This is how I'm loading via diffusers:
Safaik that one doesn't have the NaN issues
SAh, I'm not generating samples during training. I'll do that to check if there's an issue with the training, since it uses the same vae and config
S@Dr. Furkan Gözükara ok yeah, my sample looks like this. Probably the way I'm loading the vae
AHi, i have this error now when i run the kohya interface
ZWhat's the best way to fix the blurryness in this photo, upscaling doesn't seem to do anything
Fwhat do you use to laod vae?
Fyou dont use auto1111?,
Fuse my auto installer
F
Fhigh res fix usually
Ait's a kaggle problem
Fi updted kaggle notebook
Fnow works better
Fwe use working directory
Ftemp is too slow
Ffails
Athe same for the patreon?
Fyes on patreon
Fare you using latest version?
SI am using sd-scripts directly for training and then using diffusers for inference.
I’ve narrowed down the issue to bf16 mixed precision setting not loading the vae as half precision.
I’ve narrowed down the issue to bf16 mixed precision setting not loading the vae as half precision.
SNot clear to me why the vae loading doesn’t support bf16 and instead loads the full float version
Fdiffusers inference problematic
