i got memory errors on 2x RTX A6000 (48GB)
i got memory errors on 2x RTX A6000 (48GB)
Fyes you cant train dual gpu dreambooth
Fbut you can train lora
Mi see in your training log 256_img_BS_7_Fine_Tuning.txt that you trained 8x GPU dreambooth with NVIDIA RTX A6000 VRAM 48577MB Arch
Fyes but that 8x gpu was used for different trainings
Fnot as a combination 
Flora works with multiple gpu but dreambooth fails unless you get big vram gpus
Jso whats better for fine tuned image gen realism right now?
dev or krea?
pros and cons? can someone give me their observations?
dev or krea?
pros and cons? can someone give me their observations?
Fi posted grids and hopefully will make a video today
Fstill working on wan 2.2 presets
Jwhere can I see your grids, patreon?
Jand how about wan 2.2? kling 2.1 master still better?
Fi dont use kling sadly
Fgirds here https://www.patreon.com/posts/112099700
Patreon
Get more from SECourses: FLUX, Tutorials, Guides, Resources, Training, Scripts on Patreon
F
JI appreciate you taking the time to make that
JI personally don't see any differences/patterns
Jprobably going to stick to flux.dev for now
SHey guys, can confirm the tests of our dearest Dr Furkan.
Flux Krea training is amazing but needs more epoch. I trained a makeup style checkpoints with 200+ images, and this is epoch 200. With flux dev the sweetspot is often around 140/150 epochs for my trainings.
I will train 20 epochs more on this one cause I think it may be beneficial for some details. Go ahead Flux Krea is awesome, reach for high epochs !
Flux Krea training is amazing but needs more epoch. I trained a makeup style checkpoints with 200+ images, and this is epoch 200. With flux dev the sweetspot is often around 140/150 epochs for my trainings.
I will train 20 epochs more on this one cause I think it may be beneficial for some details. Go ahead Flux Krea is awesome, reach for high epochs !
Snice, whats your learning rate?
S4e-06
Fmatching with my results
Fwritten this on patroen post :d
S@cyberbol
KGlad to see this I’ll keep training
KHave you trained a Lora on the same makeup style?
JHello Dr, do you think we will be able to do something similar to dreambooth style full finetune of qwen? Is this possible?
筋I would like to ask everyone, I used 38 high-definition half-body portrait materials to fine-tune the DreamBooth training, with a configuration of 48G and 200 training rounds. Currently, the results from round 120 to round 200 are indeed becoming more similar to the characters in the material set, but I encountered two problems: 1. The test results for cartoon images are all realistic people, 2. If the character is relatively small in the image, there is only a very small probability that the face resembles the original.
JDid you caption the cartoon images with a cartoon keyword, and photorealism images with "photo" or a similar keyword?
JMaybe also introduce images where the face is more distant in your training ? Maybe you could pad the training images twice, but the second time with a big black border, so the face comes out relatively small?
Fdreambooth later i think
Ffirst lora
Fhow do you test?=
F2 : yes zoom in character
筋comfyui，Grid test
筋I relabeled the data again, added a realistic style, and plan to retrain for DB testing.
Fyou are training a game character or something?
筋I'm using the krea flux base model. I don't know if this is the reason.
筋I'm training a LoRA of a character similar to you using my own photos. However, currently, when I write the cartoon prompt, it turns into a real - life person.
筋According to your suggestion, I train in the form of DB to achieve the effect of LoRA.
Foh you train real person and want cartoon output?
筋I saw that there are cartoon - style characters in your videos.
Fyes they are hard
Fif you want cartoon flux kontext much better
FYouTubeSECourses
Black Forest Labs (BFL) publisher of FLUX models kept their promise and published the FLUX.1 Kontex DEV model today. So I didn't sleep and after doing huge research and test non-stop, I have prepared this excellent step by step tutorial that will show you how to use this amazing model. With FLUX Kontext you can edit any part of the image in any ...
Fand flux krea will make it even harder
Ffor cartoon
筋I can use kontext, and I have successfully trained my own control lora. At present, I just want to replicate the process in your tutorial, use a dataset of real - person photos to train a DreamBooth (DB) to achieve a personal model with generalization ability.
筋Thank you for your reply.
Fuse flux dev not krea for this task
Fflux krea is optimized for realism