I was running two trainings on two individuals, the saved epochs are blending faces of the two people - because my models were saved in the same output folder

Zzaiar hey quick question, im trying to train a LoRa on Flux, I have a 3060 with 12gb s...

zaiar•12/17/24, 2:17 AM

i have 32gb of RAM as well if that clarifies, any help would be greatly appreciated

very new to this

FFurkan Gözükara SECourses SwarmUi\models\Diffusion Models\a

QMOP•12/17/24, 2:18 AM

I saved every 14 Epoch - the first epoch was ready so i tested it - but it is not generating anything near the original people

QMOP•12/17/24, 2:42 AM

The second model after EPOC 28 has gone even further astray

�

🍭🎀 𝒜𝓋𝒶 𝐹𝓇𝒾𝑔𝑔 🎀🍭•12/17/24, 10:41 AM

best caption type for sdxl? training prompts in your joy caption app?

Zzaiar i have 32gb of RAM as well if that clarifies, any help would be greatly apprecia...

Furkan Gözükara SECourses•12/17/24, 11:31 AM

yes please upgrade to 64 gb and all training errors will be fixed

�🍭🎀 𝒜𝓋𝒶 𝐹𝓇𝒾𝑔𝑔 🎀🍭best caption type for sdxl? training prompts in your joy caption app?

Furkan Gözükara SECourses•12/17/24, 11:31 AM

pre alpha good

�

🍭🎀 𝒜𝓋𝒶 𝐹𝓇𝒾𝑔𝑔 🎀🍭•12/17/24, 11:32 AM

hm? it's either descriptive, rnd tags or training-prompts.

�🍭🎀 𝒜𝓋𝒶 𝐹𝓇𝒾𝑔𝑔 🎀🍭hm? it's either descriptive, rnd tags or training-prompts.

Furkan Gözükara SECourses•12/17/24, 11:33 AM

i would say it depends on your base model

Furkan Gözükara SECourses•12/17/24, 11:34 AM

make few comparison what model generates for each type

FFurkan Gözükara SECourses yes please upgrade to 64 gb and all training errors will be fixed

zaiar•12/17/24, 1:56 PM

its a little bit out of my budget right now, I just bought the 3060 to upgrade from 8 to 12gb of VRAM, and then I bought 2 more 8gb of RAM to increase from 16gb to 32. is it possible to complete training with 32gb of ram? even if it takes a long time

zaiar•12/17/24, 1:59 PM

I monitor my memory usage during training, my RAM stays at about 25-28GB of usage through-out the process, but seems like my GPU is maxing out, it'll hit 15.1gb of VRAM usage and then 2 seconds later the training crashes

zaiar•12/17/24, 1:59 PM

ill run the training and copy paste the cmd message later, not home right now

Timson•12/17/24, 3:01 PM

@zaiar just do your AI research in the cloud, it's not worth upgrading local PC to match specific requirements unless you are planing to do it as a full time occupation or you have other good uses for it (like gaming PC).
6.5$ gives you a whole work day access to a monster configuration server. It's good if you just want to experiment from time to time and requires no long term investments

Timson•12/17/24, 3:06 PM

@Furkan Gözükara SECourses what is the maximum number of dreambooth training steps you performed on flux-dev until you encountered serious model sanity degradation? I'm starting to see 6 finger hands on 3000+ steps. Can regularisation help with fine tuning model to a more complex concepts (like entire movie franchise visual style, instead of specific characters)?

Yobin•12/17/24, 5:16 PM

@Furkan Gözükara SECourses what's your opinion on doing dreambooth flux training with style then use the checkpoint dreambooth fine tuned for character lora training? I want to seperate the style and character models

Jassim•12/17/24, 5:35 PM

@Furkan Gözükara SECourses im trying to use kohya ss to finetune sdxl on runpod using your method its taking forever for step1/2 and i dont even know if its even still running or stopped

How can i do multiple people on a single finetune

Similar Threads

Similar Threads

Similar Threads