yes, i saved every 2nd epoch so i have 2, 4, 6, 8 but all of them makes this kinds of weird bugs
yes, i saved every 2nd epoch so i have 2, 4, 6, 8
but all of them makes this kinds of weird bugs
but all of them makes this kinds of weird bugs
FViT-bigG-14', 'laion2b_s39b_b160k
Fdm me your print training command
EI get this two photos when i try "photo of ohwx man" cfg 5 no adtailer
Fsomething wrong with your training possibly
Esent you the command on dm
F1 more thing
Fcan you test with
FDPM++ 2M SDE Karras
F40 steps
AI have 8 gb VRAM:)
And this settings really helped for me
And this settings really helped for me
Fnice
Fyes --medvram super important
Feven 12 gb need it for sdxl
AYeah, I compared the SDXL and SD 1.5 images and really the improvements are not worth the increased system requirements for comfortable use. I think it will be at least another year before SDXL is would be fully usable by most people.:)
Fwell i relly find SDXL superior
Mone area I realized with training on 48 gb vram and using models on 3060 12gb vram was that high quality may have been trained into the model but you can't access (recreate) it with lower vram systems. My trained models have come to life with an upgrade to 4090 24gb vram. I didn't think the difference would be so noticeable. SDXL has much improved prompt response and much higher quality. I recommend that if someone wants both high quality, flexibility and speed runpod or kaggle is the best bet. I wouldn't recommend going back to 1.5 - granted it is an expensive upgrade..
MAnother comparison: Adafactor vs Prodigy.
The Adafactor is slower to train and holds the prompt better on these lora. The lora trained with Prodigy is strong and the training is fast, (4th epoch seems to be good if it's just a face). After that it picks up other features (hair, body shape, tattoos, etc.)
So for a person I prefer Adafactor, for a style or concept I prefer Prodigy, but if you need a realistic person you can use Prodigy for that too, but you may need Clip Skip 2 or 3 for flexibility when generating person images, because Clip Skip 1 is just an image generator, nothing more.
The Adafactor is slower to train and holds the prompt better on these lora. The lora trained with Prodigy is strong and the training is fast, (4th epoch seems to be good if it's just a face). After that it picks up other features (hair, body shape, tattoos, etc.)
So for a person I prefer Adafactor, for a style or concept I prefer Prodigy, but if you need a realistic person you can use Prodigy for that too, but you may need Clip Skip 2 or 3 for flexibility when generating person images, because Clip Skip 1 is just an image generator, nothing more.
FI am preparing SOTA image captioners 
Fauto install auto run so easy
PSOTA?
Famazing script coming
FSOTA means state of the art
Pahh. thought it was a model I'd never heard of, lol. I've been looking at blip models to do captioning. but not sure which one would be optimal. it's really the main reason I haven't gotten into LoRA training on sdxl. don't want to spend hours captioning when I can get ai to do it. then all I'd have to do is go back and make sure the captions align with what I'm wanting
Fi am preparing best ones
Fblip2
Psweet. yeah, I was looking at the blip2 models on the salesforce huggingface. but sooo many models
FSOTA (The Very Best) Image Captioning Models Script For Stable Diffusion And More > https://www.patreon.com/posts/sota-very-best-90744385
Frecording a video
BI'm trying to build a system for creating family portraits where you can put your own pictures of family in it. Do you guys recommend google collab to do the renders? https://youtu.be/e2THe5hhqss?si=uz6BNk635tljw3BI
https://youtu.be/e2THe5hhqss?si=uz6BNk635tljw3BI
https://youtu.be/e2THe5hhqss?si=uz6BNk635tljw3BI
Fcolab i dont suggest
Frunpod much better
ASo, I think it makes sense for me to upgrade my GPU to fully work with SDXL. I'm used to using SD locally, so I'm not considering runpod yet, but I'll look into it just in case:).
Fi updraded rtx 3090 just for Stable diffusion
EAny tutorial on SDXL training/ creating LORA on Google Colab? I have the pro version so want to utilize that.
Fsince colab banned free i couldnt make one for it
Fi have for free kaggle and also for runpod
Fprobably kaggle version would work directly on colab
Fwith just changing folder path
Eyeah I was thinking the same. I'll try to change the folders and see if it works.
I'll post a link of the working colab here afterwards.
I'll post a link of the working colab here afterwards.
PI haven’t yet look the workflow for SDXL yet, is it possible to train dreambooth model using a 4090? Or do I have to rent a runpod for that?
Jw/ a 4090 you can train locally using DB - basically need 24gb.
AHello @Dr. Furkan Gözükara. I have used s3fs to mount aws s3 bucket onto system. I have used kohya_ss for training using the images from s3fs mounted folder which was successfull. I've mentioned the s3fs folder in additional networks path in settings of sd-webui, but lora files are not showing although they exist in folder.
MHey @Dr. Furkan Gözükara , do you have a .JSON / .TOML file for the best parameters to use in KOhya SS to train a very realistic Model ( i prefere Photon / Realistic Vision V5 )
The file i'm talking about is the one containing all these parameters .
( Alpha , fp16 .; prior_loss_weight = resolution = enable_bucket = min_bucket_reso = max_bucket_reso = .. )
is it free or is it only for patreon guys ??
The file i'm talking about is the one containing all these parameters .
( Alpha , fp16 .; prior_loss_weight = resolution = enable_bucket = min_bucket_reso = max_bucket_reso = .. )
is it free or is it only for patreon guys ??
Enot good either
Ecfg 5 feels like better results but not so flexiable, might have over trained it from the start somehow?
Ealso Adetailer gives corrupted faces
