if training on de-destilled flux model in kohya, I guess prompts will simply fail to work?
if training on de-destilled flux model in kohya, I guess prompts will simply fail to work?
Fwhat do you mean?
Fi trained on de-distilled
FMedium
Thoroughly experimented with Fine-Tuning / DreamBooth training of Flux-dev-de-distill, PixelWave v03, Verus Vision and base FLUX Dev model…
KI mean, if I use "sample prompts" in kohya ss, it will think this is Flux model, but as I have de-distilled, this needs different CGF strategy? Its like, in Comfy UI, I need different workflow, if I use de-distilled model?
Kyou probably simply did not use sample prompting when training?
Fwhat changes is
Finstead of cfg 1
Fwe use cfg 3.5
Fboth training and inference
KI see, thanks. I'll disable "sample prompting" for now, seems like not yet way to supply it to kohya ss
Kas only has "cfg scale" as an option there
Fye i dont recommend samples
KI just like to sneak-peak early on how it goes 
but I guess I can just stop training at cehckpoints, render previews manually, and resume from state if needed
but I guess I can just stop training at cehckpoints, render previews manually, and resume from state if needed
ZHello everyone! Owners of the RTX 3090 and CPUs (WITHOUT integrated graphics), has anyone managed to achieve the training speed 7.3 s/it stated in the config by @Furkan Gözükara SECourses ? Even when I don't turn anything off, I still end up using 1000GB of VRAM on Windows 11, even if absolutely no third-party programs are running. How it possible ? i got onaly 11.16 s/it with 24GB config file...
Kabout VRAM, I notice one thing. I have RTX 4090, on 24GB config, I have 6s/it; Then I swap blocks to 12GB config, and have 6.7s/it, almost no difference, but now I have lots of free VRAM and can use my PC normally alongside training. Have no idea why is that so
Kmaybe its because I have 128GB of CPU ram
Fsomething wrong
Fdid you do a fresh kohya install?
Fyou should get better than 6s actually but still not too bad
Roppions on this aws instance type https://instances.vantage.sh/aws/ec2/g6.2xlarge
The g6.2xlarge instance is in the gpu instance family with 8 vCPUs, 32.0 GiB of memory and up to 10 Gibps of bandwidth starting at $0.9776 per hour.
RThe g5.xlarge instance is in the gpu instance family with 4 vCPUs, 16.0 GiB of memory and up to 10 Gibps of bandwidth starting at $1.006 per hour.
WMy speed is about 6.5s/it
NI don’t really get it so what should I put in the fields? https://media.discordapp.net/attachments/1194438792396931174/1311703395975041074/image.png?ex=674b23fb&is=6749d27b&hm=a0a3f699f62d559dc707d0f32885528688c8be7ff59f868dbda6cf6540e28a4b&
R@Furkan Gözükara SECourses do you have intrsutions on how to install the required software on ubuntu
Flike 5x expensive
Fput 10 and 10
Fnope but you can look my scripts
Fanalyze them
Ryeah but they are aculty avaiable and you can keep your fgfiles without being charged
Fwhat do you mean
Ri find that massedcompute gpu instances are often unavaiable
Ftrue
Fyou can use runpod meanwhile
Fbut massed compute will increase hopefully soon
Falso i am talking with another provider
Rgcp seems to have some good options as well. easy to setup
Rdigital ocean has high end serve as well but it is overkill
RH100
Fye
Rcurrentl 3.50/gpu/hr
Fso expensive
Rthe gpu costs like 20k
�the kaggle, how can I remove anything from the temp folder...I only see the working folder.
Fyou cant see but it is there
F!rm -r folder path
�ah yes, I tried downloading juggernaut11, first from huggingface, didn't get it to work, used api token and all, but from civitai it worked with wget, but forget the end part with -O "filename.safetensors", so model wasn't found in supir. the filename became the download link............haha
F
Gis it possible to convert a fine tuned flux dev model into a lora?
Dyou can extract a lora
