Any tip for make the sdxl imagen generation more fast? take a lot of time generate just one pic, I
Any tip for make the sdxl imagen generation more fast? take a lot of time generate just one pic, I am using 3090 in runpod btw
Iyou still have no disk space left on your pc
Fthis could be due to selected sampler or vae? have you tested different samplers more step count?
Fcaption all images good
Fyou wont need trigger word
Fyou can use my SOTA captioners
FYouTubeSECourses
Download scripts from here : https://www.patreon.com/posts/sota-very-best-90744385
Full length Twitter video here : https://twitter.com/GozukaraFurkan/status/1712665890259447831
SOTA (The Very Best) Image Captioning Models Script For Stable Diffusion And More
1 Click install and use SOTA image captioning models on your computer. Supports 8 bi...
Full length Twitter video here : https://twitter.com/GozukaraFurkan/status/1712665890259447831
SOTA (The Very Best) Image Captioning Models Script For Stable Diffusion And More
1 Click install and use SOTA image captioning models on your computer. Supports 8 bi...
Fit works for all. concept training you mean general fine tuning?
Fyes your runpod is out of space
Fturn off pod and increase space or delete some checkpoints
Fwith
Frm -r filename
Ftensorrt
Fhopefully i will make a tutorial for that for runpod too
MI am happy now. I looked at several of Lora's profiles that used poses, concepts, clothing features. Based on that, I put together the file below. However, I also noticed two tricks:
- they set a 10x repeat regardless of the number of images
- usually for 10 epochs, so the number of images x100 is the measure in most cases.
This time, however, I used WD14 tagging on SD 1.5. I removed the labels that were valid for my concept and a few that could possibly affect the final result of the generated image (e.g. blonde hair, brown eyes), and I left one of the double labels as phone and cellphone.
And indeed, the 10th epoch model became absolutely flexible. And it even works without a keyword! I love it, and it shows how important the tagging and the repetition of images and the number of steps are.
- they set a 10x repeat regardless of the number of images
- usually for 10 epochs, so the number of images x100 is the measure in most cases.
This time, however, I used WD14 tagging on SD 1.5. I removed the labels that were valid for my concept and a few that could possibly affect the final result of the generated image (e.g. blonde hair, brown eyes), and I left one of the double labels as phone and cellphone.
And indeed, the 10th epoch model became absolutely flexible. And it even works without a keyword! I love it, and it shows how important the tagging and the repetition of images and the number of steps are.
MNot just fine tuning, but also a new keyword. I experimented with Onoff model, where you have to learn the model:
- the split screen,
- the person on the left and the person on the right are the same
- the person on the left has clothes on, the person on the right has clothes off, same body shape
- maybe even wear the same pose on both sides of the person
- while remaining flexible on face, dress, background, but background on both sides is fine if the same.
- the split screen,
- the person on the left and the person on the right are the same
- the person on the left has clothes on, the person on the right has clothes off, same body shape
- maybe even wear the same pose on both sides of the person
- while remaining flexible on face, dress, background, but background on both sides is fine if the same.
DDon't understand sir
Fye you need to understand how training happens. long thing to explain. you need to watch all tutorials 
FarXiv.org
Large text-to-image models achieved a remarkable leap in the evolution of AI, enabling high-quality and diverse synthesis of images from a given text prompt. However, these models lack the ability to mimic the appearance of subjects in a given reference set and synthesize novel renditions of them in different contexts. In this work, we present a...
Lwhen completing training with dreambooth, why do we merge the trained model with the original model used?
Fwe dont merge
Lany reason why some people merge and some do not?
Fthey could be merging to get other model style
Fbut no need to merge with base model you trained on
Ecould you tell us which models did you use to train your lest set for the costumer you said about ealier? when i tried a few models from there it wasnt very accurate. Also, as discussed here before, why use captions when you use regularlazation images?
D@Dr. Furkan Gözükara Sir are we talking about this 
Fi use SDXL 1.0 base
Fcaption didnt affect much
Fi think most important part is training images quality and selection
Eoh i meant what captioning model sorry
Fit is a long thing to explain here
Fi used blip2
Fyou know i shared on patreon
Eactually the set you helped me in private (eliminated a few images) - turned out my best model ever (out of around 10 made)
Dok ok , i will just go through them.. i thought you are speaking about the captioning.. NVM , will look into your tutorials ..
Ei understand its blip2 but when opening the captioning webui there's a really huge list of models thats the one im talking about
Fas you watch more you will understand more haha i can also make private session and explain you if you are interested in professionally
i can also make private session and explain you if you are interested in professionallyEbut if you say captioning didnt matter much it doesnt matter
Fblip2 is stand alone
DI am now seeing your controlnet tutorial sir. Going ot to take the nexct step into controlnet , and then dreambooth as you suggested
F
DYouTubeSECourses
I show how to install Automatic1111 Web UI & ControlNet extension installation from scratch in this video. Moreover I show how to make amazing QR codes and inpainting and out painting of ControlNet which are very similar to Photoshop generative fill and Midjourney zoom out. Furthermore, I explain and show what are Canny, Depth, Normal, OpenPose...
EI've actually tested so many models but forgot to document (or rather was in a hurry from excitment) - is there a way to take a checkpoint and somehow check it to see if i used captions to train it?
L@Dr. Furkan Gözükara training with dreambooth with lion, can I use it without using classification images?
Lfor this setting i set it to 0 or leave it to 5?
Fdont use it
Fuse my class images
For generate from txt2img tab
Fmanually
Lfor 30 images how many images in my classification dataset and how much "class images per instance images" should i set it to?