i trained 6 models like 2 days ago
i trained 6 models like 2 days ago
Fworked perfect with my config.
GI just solved a problem I had. When CompyUI connected vae to vae decoder through the vae loader node, the image was created normally. I was using the config file you shared with Patreon.
Fmakes sense because training werent broken 
GI trained the model using your config file, but when I didn't connect the VAE loader, the images came out the same as before. Is it possible that the model didn't train properly? I used the config file you shared on Patreon and trained with normalized images.
KSpooky
Fuse models in auto1111 and if they work directly then model trained accurately
M@Furkan Gözükara SECourses have you tried trainings with sdxl noise offset of 0.0357 and the scheduled huber loss settings
M@Furkan Gözükara SECourses when you get some time can you try this prompt out on your training?
Postive - cinematic photo Portrait of ohwx man as super Mario, red hat, fantasy, highly detailed, digital painting, artstation, concept art, sharp focus, illustration, art by Tony Sart and artgerm and randy vargas. 35mm photograph, film, bokeh, professional, 4k, highly detailed
Negative - drawing, painting, crayon, sketch, graphite, impressionist, noisy, blurry, soft, deformed, ugly, 3d render
DPM ++2sa Karras
1024x1024
CFG - 7
can you check if you get good likness
Postive - cinematic photo Portrait of ohwx man as super Mario, red hat, fantasy, highly detailed, digital painting, artstation, concept art, sharp focus, illustration, art by Tony Sart and artgerm and randy vargas. 35mm photograph, film, bokeh, professional, 4k, highly detailed
Negative - drawing, painting, crayon, sketch, graphite, impressionist, noisy, blurry, soft, deformed, ugly, 3d render
DPM ++2sa Karras
1024x1024
CFG - 7
can you check if you get good likness
TAfter 30 training and 5 dataset retake to have the best one, i finally achieve the - p e r f e c t - result thanks to @Furkan Gözükara SECourses however, mid-distance or distance shot are Totally destroyed, my dataset is perfect, used cinema grade camera, used cropper and resizer, tried mask training, adetailer, changing the number of step only for face, changing the denoising strengh, i can only make p r e f e c t close-up, selfie or whatever, but when the subject is in midrange like your picture in civtai, or fullbody, it's totally crap, even the face is not really "human" more like destroyed, any idea ? thanks for everything (also tried with and without reg)
Mare you training on realvis4 and using the model on a different checkpoint?
Ttraing on realvis4, using on realvis4
Ti have to admit i spent the last weeks understanding every little things possible and even myself cannot recognize a (fake) picture from myself, ,even my family, but im limited to close-up
Mand all your images are cropped at 1024x1024?
Tok i understood, to be possible to inpaint in a far away face, you need to Hires a lot, so the Adetailer can actually see the face to inpaint
Muse adetailer with correct settings, and with hires fix use 4x ultrasharp or any other good upscalers
Tmy problem was that i wasnt upscalling enough, 1.25 or 1.5 is not good enough for the adetailer to work properly
Tbut 2x upscaling is good, now i have the (body proportion) problem but its easier to fix than not be able to inpaint
Myeah i usually upscale with 1.75x to 2x with 60 steps on adetailer
Fsome prompts will overwrite your likeliness
Flike super Mario
Fbecause the model will assosicate it with a certain man
Fsuper Mario - original dataset
For super man
Fit is like ohwx man as brad pitt same
Ffor mid and distant shots i use adetailer and it works
Fyou wont get good face with those without adetailer
Fbecause the model inherent training rsolution is too low
Mi noticed training a little bit of the text encoder 2 helps with that
M@Furkan Gözükara SECourses what is mulit line captioning? the wildcard? what does it do and how do you use it
K@Furkan Gözükara SECourses do you have any guidelines/guide on how to pick images? as in what type of images should be picked for training?
FNice tips
FJust write a prompt each line
FIt will randomly use during training each one
FI explain in tutorials but I don't have a specific tutorial yet
FI am composing a good dataset to make a dedicated turorial
Ti tried different dataset style, always perfect quality with different clothing and background
i found out that it is not the best idea to add different facial expression since when using a prompt like "surprised"
if the model use an already surprised expression from the dataset it would be SUPER exaggerated and very weird, even using moderate prompt like "slightly"
I suppose that it would not be an issue if every picture from dataset were captioned, but i havnt found a good tutorial about dataset/captioning and method to use
Also i found that not be able to modify face is a pretty sad, adding a cyborg eye, or facepaint, makup, blood, dirt whatever, tried a lot of prompt even in the adetailer face parameter
I also found out that if your dataset is always with a sharp focused talent and blurry backround, whatever prompt you use, it will always be the same in Stablediffusion, always a sharp focused person, but blury background (i may doing something wrong), i feel like the "from single text file" method is pretty good for something quick, but if you want to achieve God level of training, you need to do the Super anoying long version of data training method
i found out that it is not the best idea to add different facial expression since when using a prompt like "surprised"
if the model use an already surprised expression from the dataset it would be SUPER exaggerated and very weird, even using moderate prompt like "slightly"
I suppose that it would not be an issue if every picture from dataset were captioned, but i havnt found a good tutorial about dataset/captioning and method to use
Also i found that not be able to modify face is a pretty sad, adding a cyborg eye, or facepaint, makup, blood, dirt whatever, tried a lot of prompt even in the adetailer face parameter
I also found out that if your dataset is always with a sharp focused talent and blurry backround, whatever prompt you use, it will always be the same in Stablediffusion, always a sharp focused person, but blury background (i may doing something wrong), i feel like the "from single text file" method is pretty good for something quick, but if you want to achieve God level of training, you need to do the Super anoying long version of data training method
SDoes anyone have some good results with SDXL trainings and could possibly share results? Mainly for close-ups / portraits.
Tried 1.5 but not too great for me..
Tried 1.5 but not too great for me..
Dwhat do you mean? theres hundreds of trained good stuff on civitai
SYeah of course but preferably from people here with advice and that can share training parameters
SCivitai is full of good stuff ofc
Tyou just follow the onetrainer tutorial from our lord Dr.
1.5 and SDXL works perfect for portrait or close-up, even midrange
The most important is dataset, high quality of 20-25 pictures with everything focused, the most important is eyes/iris
Don't do crazy angle but not only front face picture, add some slightly left/right/pan/tilted angle so you can use full potential of Stablediffusion generation
Use Realvisx4 sdxl model first so you can learn from a model that dsnt need a lot of tweeks to work great, if you want to do NSFW i cannot really help you since i do not produce them, but i assume that using Reg images will break a lot of the capacity of a model to do such generation,
You also gonna need comfy-UI and learn how to use it since you can make different layers of generation to be able to add texture in the face easily (makeup,modification and such) i suppose there is a method to do it without comfy, didn't found it yet.
If you follow Dr. tutorial, copy EVERYTHING and use a GOOD dataset, you will be able to produce even more quality than him (since he is using a medium quality dataset)
1.5 and SDXL works perfect for portrait or close-up, even midrange
The most important is dataset, high quality of 20-25 pictures with everything focused, the most important is eyes/iris
Don't do crazy angle but not only front face picture, add some slightly left/right/pan/tilted angle so you can use full potential of Stablediffusion generation
Use Realvisx4 sdxl model first so you can learn from a model that dsnt need a lot of tweeks to work great, if you want to do NSFW i cannot really help you since i do not produce them, but i assume that using Reg images will break a lot of the capacity of a model to do such generation,
You also gonna need comfy-UI and learn how to use it since you can make different layers of generation to be able to add texture in the face easily (makeup,modification and such) i suppose there is a method to do it without comfy, didn't found it yet.
If you follow Dr. tutorial, copy EVERYTHING and use a GOOD dataset, you will be able to produce even more quality than him (since he is using a medium quality dataset)
Mrunpod prices decreased a lot for 6000ada and other gpus, looks like they'll try to compete with massed compute
M@Furkan Gözükara SECourses upgraded to new xformers 0.27 and torch, getting way better results and smoother results with xformers and training now on kohya
SI'm looking for an AI developer to develop my project.
If u are interested, contact me.
If u are interested, contact me.
Fyou got some very accurate findings
Ftrue i see A6000 dropped from 0.69 to 0.49 but still we get from 0.31 cent on massed compute with our coupon
Fyes xformers newer versions are way better than older ones
SThanks a lot for this! Really interesting I appreciate it.
I’ve done a SDXL training with a pretty good dataset I think, on Realvisxl 4, no captions. But the quality wasn’t that great, and the face resemblance was even worse.
https://discord.com/channels/772774097734074388/1261270101244641302/1261336695110897766
I use Comfy as my main tool but the training itself must be a little bit better I feel like, maybe I’m still doing something wrong
I’ve done a SDXL training with a pretty good dataset I think, on Realvisxl 4, no captions. But the quality wasn’t that great, and the face resemblance was even worse.
https://discord.com/channels/772774097734074388/1261270101244641302/1261336695110897766
I use Comfy as my main tool but the training itself must be a little bit better I feel like, maybe I’m still doing something wrong
Tdon't forget that you need to use the good vae if not baked-in, Hires 1.5/1.7, adetailer capped on resolution 1024x1024 with a denoising strengh between 0.35 to 0.60,
TAlso i did an update on my stable diffusion and i couldnt anymore use my training face, , everything worked, but not my trained model, suddently turned bad for no reason, i had to reinstall a fresh sb, dunno why but some ppl had the same pb as me, so maybe you have an adetailer version or something that went wrong