I found it! SwarmUI Windows Installer.bat
I found it! SwarmUI Windows Installer.bat
Fyep
F
KQuestion for all!
How long does it take to generate an image in Forge on your GPU?
Resolution: 1920x1080
Model: flux1-schnell-bnb-nf4.safetensors
VAE: ae.safetensors
Steps: 20
Sampler/Scheduler: Euler/Simple
It feels like my 1 min. 8.9 sec is way to long!?
How long does it take to generate an image in Forge on your GPU?
Resolution: 1920x1080
Model: flux1-schnell-bnb-nf4.safetensors
VAE: ae.safetensors
Steps: 20
Sampler/Scheduler: Euler/Simple
It feels like my 1 min. 8.9 sec is way to long!?
KPython 3.10.6
Pytorch version: 2.1.2+cu121
Pytorch version: 2.1.2+cu121
KI seem to be running an old version. I wonder if newer ones are faster?
Fyou are missing most important variable
Fi am finally testing accurate microsfot c++ build tools version
![Dazzastrous [4090]](https://cdn.discordapp.com/avatars/871389071686135838/c4a5f4d81d9eeb18d856295d692e2c07.webp?size=40)
Dplease?
Dlol
KWhich one?
Doh Doc the irony u putting accurate and then Microsfot
Dhehhe
Fyour gpu model :d
KTesting "Forge with CUDA 12.1 + Pytorch 2.3.1" and "Forge with CUDA 12.4 + Pytorch 2.4" to see what the difference are
![Dazzastrous [4090]](https://cdn.discordapp.com/avatars/871389071686135838/c4a5f4d81d9eeb18d856295d692e2c07.webp?size=16)
Fye i hate this c++ thingie
Fit is time to finalize it :d
Dyeh me too
Fi will make a video for this
KDamn
D30.2 secs
KYeah well... 4090 sounds right
Ffaster internet so useful
Fswarmui has --fast
Ffor rtx 4090
Freduces flux to like 10-12 second
F20 steps
D
Doink oink
Dflux is a bit heavy on the boka , SD more detailed backgrounds
Dmore detail with sd
D
Dlike the colours in this
Fhaha i made a thread on all nvidia githubs
Fhttps://github.com/NVIDIA/CUDALibrarySamples/issues/217
https://github.com/NVIDIA/cccl/issues/2383
https://github.com/NVIDIA/cutlass/issues/1793
https://github.com/NVIDIA/cuda-samples/issues/297
https://github.com/NVIDIA/TensorRT/issues/4114
https://github.com/NVIDIA/cuda-python/issues/88
https://github.com/NVIDIA/cudnn-frontend/issues/106
lol they have 500 repos
https://github.com/NVIDIA/cccl/issues/2383
https://github.com/NVIDIA/cutlass/issues/1793
https://github.com/NVIDIA/cuda-samples/issues/297
https://github.com/NVIDIA/TensorRT/issues/4114
https://github.com/NVIDIA/cuda-python/issues/88
https://github.com/NVIDIA/cudnn-frontend/issues/106
lol they have 500 repos
FPatreon
Get more from SECourses: Tutorials, Guides, Resources, Training, FLUX, MidJourney, Voice Clone, TTS, ChatGPT, GPT, LLM, Scripts by Furkan Gözü on Patreon
Fi will update article testing still
SThe worst thing about running the kohya flux training update. "Haha you fewl, you installed torch 2.5. I will now downgrade you to 2.4 and complain about dependencies needing 2.5" And then the installer runs off like a 1920's villain and ties a damsel to train tracks
Fi overwrite gui file to prevent it
to prevent it
Xwill give this a try now
D@Dr. Furkan Gözükara One question, I'm trying to train flux with a big dataset of 1500 images, with the sugested learning rate is too slow to converge and get any results, I'm trying now with 4e-4 lr and I getting results much faster and until now looks good, do you think is too aggressive lr? may be I can start with this high lr and then finalize the training whith a lower lr to learn more finer details but without having to wait 15 days of training
Fwell the LR bigger certainly it will learn faster but it is like you are slicing the cake with a bigger knife so the minor detailed cuts becomes impossible
Fyou can speed up training with multiple GPU
Falso bigger batch size requires bigger LR
Dyou right, but by now I doing everything on my local computer, do you think that finalize the training with a lower lr can improve the final results, startin with a high learning rate and when is aready on track I can reduce the lr, do you think that lowerimg the lr at the end can be beneficial?
Xjust started training 
FMan I need 48 gb rtx 5090 comon Nvidia 
FI also need a mb that can perfectly run 3 gpu at the same time so I can test all 
F3060 3090 and 5090
