If You Could Help Me Make It Or Know Somebody Who Could Make It???.
If You Could Help Me Make It Or Know Somebody Who Could Make It???.
Fi doubt that you can find anyone with this way
FThere Must Be Some Way.
FLensa Made It Possible.
FRemini Followed Suite, There Is Something They Know That I Don't. What Possibly It Could Be.
Fof course it is possible
Fit just cost money and time to get it done
MAnyone else got weird workflows?
I find the best likeness I can get to my subject is if I train using automatic1111 Dreambooth then use Kohya to extract a LoRA from it. It’s nearly always way more accurate than training a Lora by itself in Kohya. I can also generate a bunch of checkpoints and extract all of them to find the one with the least overfitting.
I find the best likeness I can get to my subject is if I train using automatic1111 Dreambooth then use Kohya to extract a LoRA from it. It’s nearly always way more accurate than training a Lora by itself in Kohya. I can also generate a bunch of checkpoints and extract all of them to find the one with the least overfitting.
Fi only used web ui dreambooth extensin and textual inversion so far
ZTrained a model for like 6 hours today in auto1111 dreambooth, the sample images looked just like my subject, but when i go to txt2image, and put in the lora name, it looks nothing like my subject
Fsomething must be off not making sense
Fi am still processing new video
Zyeah it's weird, the instance name like photo of ohwx isn't used unless i convert it to a ckpt right? i should just be able to put in the lora tag <> and it should work
Fi think now it supports lora activation
Fare you acivating your loras?
Zhow do i activate it?
Fi didnt test newest version yet
Fyou need to type lora file name
Zi mean i'm going into the additional networks, going to lora tab, and clicking on the pt file
Fif you generate ckpt
Fthen it is ofc activated
Fwith ohwx man word
Zso it's putting the <lora model name> for txt2img, but like i said, it looks nothing like my subject, even when i use the 1.5-pruned that it was trained on
Zi checked lora and lora extended
Zdid about 300 epochs
Fi will make a new lora video hopefully soon
Zit's just strange cause every 25 epochs it was printing out images that looked like my subject, than it finishes, then looks nothing like my model
Zi'll print out the settings i used here in a sec as soon as my venv folder regenerates
F
ZGot this weird error while trying to convert my lora to a checkpoint Exception compiling checkpoint: The size of tensor a (4) must match the size of tensor b (320) at non-singleton dimension 1
MI also have a weird one where if I use the Lora I trained in comfyui it looks perfect. If I use it in automatic1111 it displays some gibberish with exactly the same prompt.
Zinteresting
Zmaybe i'll generate an image of my lora model on comfy and see what i get
MYeah try it. I still haven’t worked out why it is doing it yet but my head was aching so will look tomorrow.
Zhow do i load a lora model on comfyui
A@Furkan Gözükara SECourses It may be helpful to you if you create channels for the different GPU Cards people have, like 20series, 30series, 40 series, 16series, that way people can confirm if your updates worked on their graphics cards, and also people with those graphic cards can go see if they should update or wait...
A
MBring it in after the model but before clip. I can’t remember off the top of my head but I think the node is under loader and then Lora.
Zyeah i don't get it, i make a model and when i activate it, it's completely different
Zwhen are they going to fix this program
MSo it worked normally?
M
Mthat’s where you put the lora.
Fye could be. the problem is i dont have card to test :/
MI also think you need a trigger word for Lora depending on how well or even over trained it is.
MIf it’s way over trained probably no trigger word is necessary.
MI recommend taking a brief look at Kohya and Comfyui. Both are excellent for different things and compliment automatic1111 brilliantly. The fact I have different outputs with the same prompts, models, Lora and sampler is also interesting. Would love to see you give comfyui a go with your images. And maybe try training them using Kohya too.
MJust to see how you found it.
Fye i should cover
Fbut first i plan a comprehensive lora training of web ui
Moh no need for a video. Just would love to hear what you think.