Do they still? It mentions cropping 512 still down to
Do they still? It mentions cropping 512 still down to
AWeird, it turns my images sideways and crops into the face for some reaosn, no idea why
AAny idea when why when I loaded the 24gb text json it doesn't autofill the additional params? usually this is longer I think
Dcan you run animatediff on 3080 10 gb ?
Mwould you say its better to do dreambooth training via the extension on automatic1111 or do it on khoya?
Fyou can delete that you dont need
Fi dont know didnt test recently
Fi see auto filled
Fi am yet to explore animate diff
Dit produced a bujnch of pics which were related but not worthy to be made into gif , if thats the output.. have to explore. The prompr length also seems to matter..
Wbest hyper parametrs , its about these you posted on patreon 24GB_TextEncoder kohya ? , and how many images did you select for training, and repeats while using class images
Wif im gonna use 3 4 gpus , i also need consider about for choosing right number of repeats , as u said , if i have two gpus need to devide it on number of gpus
AVery inspiring, have you considered checkpoint training with EveryDream2 instead of LoRa...and watch these same parameters? Just wondering whether this could lead to even better results. Although I agree with you, the quick turnaround of LoRa training is great from a feedback cycle point of view.
Fi used 14 images
Fmultiple gpu means increased batch size
Fso 2 gpu will make it batch size 2. so you should halve the number of steps
Fyes that training done with 24gb config on patreon
F24GB_TextEncoder.json
WThanks , Repeats should be around 160 for sdxl , I remember for 1.5 it’s easy to overtrain and 25-40 was okay
Dthis time , the image creation needed very little effort to bring the face output. Using correct , HQ face images , makes lora training much easier.
ADo we use extract faces or resizer?
Ais this correct? I just want to do 1024x1024
MI find the big model training unnecessary. I'd only do it if I really wanted to make a new model, but that would require millions of photos, and I only do DreamBooth very rarely (for example for a pose that no model can do by default, like a full undereye). For everything else, I have Lora, which is fast, perfect for me, and the quality is absolutely comparable to a DB model, plus I can put it on top of any model, mix it with other Lora.
MNeed only one DATA array only for all size what you need, example:
DATA = [
((512, 512), r"G:\man dataset\new_cropped\512x512"),
((768, 768), r"G:\man dataset\new_cropped\768x768"),
((1024, 1024), r"G:\man dataset\new_cropped\1024x1024")
]ACan I just delete the second data = [] one?
Myes
Mlast comma no need
Aits being strange
Acropping and rotating
Aand also its 512?
MYes, it's 512. Have you used the Yolo script before?
Anope
MFirst is the Yolo script, it cuts the person. Then select the best images by hand or with CompreFace and then use this cropping tool.
Awhats extract_faces for?
MThere is a Yolo script under 30th october update:
https://www.patreon.com/posts/sota-subject-and-88391247
https://www.patreon.com/posts/sota-subject-and-88391247
M"Yolo v7 based cropper script updated and now works much better."
Mor Auto cropper (that's same)
MUnder this section found newer version on same patreon link:
How To Use Subject Cropper
How To Use Subject Cropper
MI use (1536, 1536) aspect ratio only, because need square crop later and this size is largest. No need other ratios from there
Mcropper_v7.py, line 13
MOnce you're done cutting and sorting, then use the resizer and it will cut it out really well.
MHi! I have some issues with Dreambooth Kohya, I tried to start learning but it stopped because of ‘ Doesn’t exist img folder’. I upload train dataset in workspace and pointed in dreambooth. It has happened on Runpod
ASo I have the yolo v7 in its completely own folder?
MYes, create a different folder for yolo script
Mand after install change cropper_v7.py constants, then go run_cropper.bat
Min cropper_v7.py comment second ASPECT_RATIOS line, need first one only.
Athis look good?
MYes, and if no need YOLO (tagged) images, set to False last constant
A
MI like to put the cropped images and the original images in separate folders, for example E:\woman dataset\original and E:\woman dataset\cropped