Software Engineering Courses (SECourses)•3y ago

I got this bad resaults in dreambooth training, did 8000 steps, 24 images, 42 repeats, 8 epochs Why

I got this bad resaults in dreambooth training, did 8000 steps, 24 images, 42 repeats, 8 epochs
Why did the face come out liek this ? (happens in all generations)

Sandstormleader•10/10/23, 9:01 PM

even on closer up renders?

ElisoliOP•10/10/23, 9:01 PM

im using Adetailer for face, but ill try close up

ElisoliOP•10/10/23, 9:02 PM

yes even on closeups

Sandstormleader•10/10/23, 9:02 PM

what is your cfg at?

ElisoliOP•10/10/23, 9:03 PM

Sandstormleader•10/10/23, 9:03 PM

try 3 or 4

Sandstormleader•10/10/23, 9:04 PM

i never used auto detailer so not sure on that but sometimes cfg needs to be lower

EElisoli I got this bad resaults in dreambooth training, did 8000 steps, 24 images, 42 re...

Furkan Gözükara SECourses•10/10/23, 9:04 PM

did you try previous epochs?

Furkan Gözükara SECourses•10/10/23, 9:05 PM

i am working on an auto captioner best one

ElisoliOP•10/10/23, 9:05 PM

yes, i saved every 2nd epoch so i have 2, 4, 6, 8
but all of them makes this kinds of weird bugs

Furkan Gözükara SECourses•10/10/23, 9:05 PM

ViT-bigG-14', 'laion2b_s39b_b160k

EElisoli yes, i saved every 2nd epoch so i have 2, 4, 6, 8 but all of them makes this kin...

Furkan Gözükara SECourses•10/10/23, 9:05 PM

dm me your print training command

ElisoliOP•10/10/23, 9:16 PM

I get this two photos when i try "photo of ohwx man" cfg 5 no adtailer

EElisoli I get this two photos when i try "photo of ohwx man" cfg 5 no adtailer

Furkan Gözükara SECourses•10/10/23, 9:17 PM

something wrong with your training possibly

ElisoliOP•10/10/23, 9:18 PM

sent you the command on dm

EElisoli sent you the command on dm

Furkan Gözükara SECourses•10/10/23, 9:18 PM

1 more thing

Furkan Gözükara SECourses•10/10/23, 9:18 PM

can you test with

Furkan Gözükara SECourses•10/10/23, 9:19 PM

DPM++ 2M SDE Karras

Furkan Gözükara SECourses•10/10/23, 9:19 PM

40 steps

Furkan Gözükara SECourses•10/10/23, 9:59 PM

https://twitter.com/GozukaraFurkan/status/1711863290194702506

FFurkan Gözükara SECourses it has only 6 gb vram

aelvelis•10/10/23, 10:07 PM

I have 8 gb VRAM:)
And this settings really helped for me

Aaelvelis I have 8 gb VRAM:) And this settings really helped for me

Furkan Gözükara SECourses•10/10/23, 10:10 PM

nice

Furkan Gözükara SECourses•10/10/23, 10:10 PM

yes --medvram super important

Furkan Gözükara SECourses•10/10/23, 10:10 PM

even 12 gb need it for sdxl

FFurkan Gözükara SECourses even 12 gb need it for sdxl

aelvelis•10/10/23, 10:16 PM

Yeah, 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.:)

Aaelvelis Yeah, I compared the SDXL and SD 1.5 images and really the improvements are not ...

Furkan Gözükara SECourses•10/10/23, 10:26 PM

well i relly find SDXL superior

Aaelvelis Yeah, I compared the SDXL and SD 1.5 images and really the improvements are not ...

mtnmecca•10/10/23, 10:29 PM

one 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..

mikemenders•10/10/23, 10:34 PM

Another 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.

Furkan Gözükara SECourses•10/10/23, 10:37 PM

I am preparing SOTA image captioners

Furkan Gözükara SECourses•10/10/23, 10:37 PM

auto install auto run so easy

picturesonpictures•10/10/23, 11:13 PM

SOTA?

Furkan Gözükara SECourses•10/10/23, 11:16 PM

amazing script coming

Furkan Gözükara SECourses•10/10/23, 11:16 PM

SOTA means state of the art

picturesonpictures•10/10/23, 11:22 PM

ahh. 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

Ppicturesonpictures ahh. thought it was a model I'd never heard of, lol. I've been looking at blip m...

Furkan Gözükara SECourses•10/10/23, 11:28 PM

i am preparing best ones

Furkan Gözükara SECourses•10/10/23, 11:28 PM

blip2

picturesonpictures•10/10/23, 11:29 PM

sweet. yeah, I was looking at the blip2 models on the salesforce huggingface. but sooo many models

Furkan Gözükara SECourses•10/10/23, 11:45 PM

SOTA (The Very Best) Image Captioning Models Script For Stable Diffusion And More > https://www.patreon.com/posts/sota-very-best-90744385

Furkan Gözükara SECourses•10/10/23, 11:45 PM

recording a video

BoneChef•10/11/23, 12:28 AM

I'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

BBoneChef I'm trying to build a system for creating family portraits where you can put you...

Furkan Gözükara SECourses•10/11/23, 12:41 AM

colab i dont suggest

Furkan Gözükara SECourses•10/11/23, 12:41 AM

runpod much better

Mmtnmecca one area I realized with training on 48 gb vram and using models on 3060 12gb vr...

aelvelis•10/11/23, 12:46 AM

So, 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:).

Aaelvelis So, I think it makes sense for me to upgrade my GPU to fully work with SDXL. I'm...

Furkan Gözükara SECourses•10/11/23, 12:49 AM

yes

Furkan Gözükara SECourses•10/11/23, 12:49 AM

i updraded rtx 3090 just for Stable diffusion

Furkan Gözükara SECourses•10/11/23, 1:05 AM

https://youtu.be/V8iDW8iprqU

YouTubeSECourses

Best Image Captioning Models Scripts 1 Click Install And Use For ...