is the only point of the num class instancers per image or regularization images is to preserve the
is the only point of the num class instancers per image or regularization images is to preserve the class?
Zso if it overfits, if i type woman, it will only show the woman in my dataset, and no other women?
Zcan you explain doc?
Z@Furkan Gözükara SECourses
Fnot only that
Fit also prevents overtraining
Fif you train very long, even when you dont type anything you get your trained subject
Fbecause unet becomes cooked
Zovertraining is when the images start looking warped?
Fusually
TI have a question, when you use the "ohwx man" and the <lora> to generate an image, are you still increasing the weights using ctrl+up e.g. (ohwx man:1.5) or are you keeping it at 1.1?
2nd question; Do we need both [the lora and the ohwx man] everytime we want to call our model?
2nd question; Do we need both [the lora and the ohwx man] everytime we want to call our model?
Fi usually test this because sometimes 1.1 or 1.2 works better
Fso you need to test
F2nd : yes
Tthank you!
I am seeing great results with 1.3 but gets worse higher than that
can you increase the weights on the <lora>? what effect would that have?
I am seeing great results with 1.3 but gets worse higher than that
can you increase the weights on the <lora>? what effect would that have?
Factually i didnt test increased weight of lora but you can test easily with sr prompt
TAlright, thanks again!
Zwhat's the best way to impose someones face using controlnet?
Zi like codeformer but it seems to change the face a little too much, should i lower the weight? or try gfpgan
IAhhh! and a good teacher. Because good teachers explain things as if they were learning the first time.
It is not easy to explain things as if you were explaining them to yourself, when you are already an expert in the field.
But for students it is the best way for them to learn.
And in your tutorials you can see this way of teaching
It is not easy to explain things as if you were explaining them to yourself, when you are already an expert in the field.
But for students it is the best way for them to learn.
And in your tutorials you can see this way of teaching
FThanks agreed
Ro/
Rany alternatives to kohya for LoRa training?
Ror anyone able to help me with my 4090 just getting 3.5it/s?
Fkohya is slow
Fdreambooth extension of web ui much better
Falso better quality
R
Fhave you seen my videos?
Ri think i just watched one 
Ron kohya
Rbut can i do LoRa in dreambooth extension of automatic1111?
Ralso is 16it/s in line with what to expect from a single 4090 in automatic1111?
A@Furkan Gözükara SECourses hello. I just watched your last video. Amazing stuff. One question. You said you trained The model based of The settings in that video but then you said you generated The images using realistic vision. I dont understand how that works.. Or did you train The images on realistic, because in The video you use The standard model..
Fyes you can do
F16 it is low
FYouTubeSECourses
Discord: https://bit.ly/SECoursesDiscord. RTX 3090 review along with RTX 3060. Stable Diffusion, OpenAI's Whisper, DaVinci Resolve and FFmpeg benchmarked. If I have been of assistance to you and you would like to show your support for my work, please consider becoming a patron on 
https://www.patreon.com/SECourses
Summary And Conclusions PDF
...
https://www.patreon.com/SECoursesSummary And Conclusions PDF
...Fye i used realistic vision with same configuration. so classification images and used model was different only
Fanyone reading can you please upvote i appreciate it : https://www.reddit.com/r/StableDiffusion/comments/137unbi/novel_how_to_find_best_stable_diffusion_any_ai/
Fwhat video i should i make next
Rsteering, biased
Rcan you passivly train with lora captions? like using the captions of the images which dont all represent the same object but the same theme make that theme bleed into the composition even the lora was focused on a very different object.
Fsure you can train with captions. i prefer captions only if fine tuning model
Rna iam not about training with captions but passivly train the lora and composition to be biased
Fwhat you mean by that
Rlets say you train a lora on car rims, just rims. but u use mainly pictures of cars with rims to train the lora and have the captions be very detailed about the cars to train the lora about the rims to have it focus on the rim, but will the cars bleed into a composition later if you only want to show a stack of rims
Rcan you use to passivly train the model to be more defined in regards to cars?
Rso basicly like priming LLM
Fpassive training is unet
Fso as you do unet training, the unet will likely to output your trained images
AI found this tutorial (https://www.youtube.com/watch?v=rlfhv0gRAF4) and this extension (https://github.com/CiaraStrawberry/TemporalKit) that makes the face replacement a lot quicker. I haven't gotten perfect results but the process is simpler