Software Engineering Courses (SECourses)•2y ago

I tried onetrainer with sd3 but the results were terrible, the model gets fried and no people like

I tried onetrainer with sd3 but the results were terrible, the model gets fried and no people likeness at all, the resulting images were terrible full of artifacts

Ddsienra I tried onetrainer with sd3 but the results were terrible, the model gets fried...

Furkan Gözükara SECourses•7/27/24, 12:29 AM

it needs research

Furkan Gözükara SECourses•7/27/24, 12:29 AM

i plan to do but busy got 2 client work right now that i have to finish :d

dsienraOP•7/27/24, 12:29 AM

the loss looks much better at 1024x1024, in onetrainer the loss graph is all over the place, I think there is a bug in onetrainer and sd3 I'm using the default presets on both onetrainer and kohya_ss, at this moment kohya_ss is much better on sd3, my 512x512 test looked a little undertrained but quite good

Furkan Gözükara SECourses•7/27/24, 12:30 AM

loss rates were never meanginful for me

Furkan Gözükara SECourses•7/27/24, 12:30 AM

when training SD 1.5 or SDXL

Théo•7/27/24, 12:30 AM

guys i have a big problem that started today and i don't know why, my comfyui is very laggy and the UI is super small, any idea ?

Théo•7/27/24, 12:30 AM

i feel like the resolution of the software got altered or something
edit : ok my chrome decided to put comfy ui in 15% scale resolution, super weird, but it fixed it to put it back to 100%

TThéo i feel like the resolution of the software got altered or something edit : ok my...

Furkan Gözükara SECourses•7/27/24, 12:32 AM

i was gonan tell this

FFurkan Gözükara SECourses loss rates were never meanginful for me

dsienraOP•7/27/24, 12:37 AM

yes may be, but I was able to see some correlation on loss graphs, they tend to get lower over time during training, on onetriner and sd3 the graph looks ridiculous, is a mess goes up and down and the model gets destroyed very quickly

dsienraOP•7/27/24, 12:39 AM

kohya_ss on the sd3 branch is very usable, it need some tweaks but is getting there

Ddsienra kohya_ss on the sd3 branch is very usable, it need some tweaks but is getting th...

Furkan Gözükara SECourses•7/27/24, 12:39 AM

i am still waiting him to fix

Furkan Gözükara SECourses•7/27/24, 12:50 AM

Fabricatedgirls•7/27/24, 5:28 AM

Hey Dr. Furkan do you have a good article on captioning/training photographic styles in SDXL?

Fabricatedgirls•7/27/24, 5:28 AM

like for types of lenses or lighting, stuff like that

Ddsienra I'm testing a SD3 full finetune with included kohya_ss SD3 preset, and as you ca...

protector131090•7/27/24, 7:26 AM

just change the res. i trained it many times. quality is super good. and it trains 3x times faster than xl.

FFabricatedgirls Hey Dr. Furkan do you have a good article on captioning/training photographic st...

Furkan Gözükara SECourses•7/27/24, 7:56 AM

not yet sadly

Furkan Gözükara SECourses•7/27/24, 8:03 AM

automatic1111 new version published : https://github.com/AUTOMATIC1111/stable-diffusion-webui/releases/tag/v1.10.0

GitHub

Release 1.10.0 · AUTOMATIC1111/stable-diffusion-webui

Features:

A lot of performance improvements (see below in Performance section)
Stable Diffusion 3 support (#16030, #16164, #16212)

Recommended Euler sampler; DDIM and other timestamp samplers cur...