sorry it was a clickbait video..
sorry it was a clickbait video..
Fyep
FGitHub
Currently create_SUPIR_model inside util.py loads everything into RAM first with full precision Thus it is using more than 29 GB RAM I think model half param can be given to it as well Or perhaps s...
Pi think you can use fp8
Pits use same stable diffusion engine that a1111
Pi am busy with work jobs i cant try it yet
Pyou can send that A1111 PR to they or we can be the first to do this 
Pi think they will be very lazy to implement these optimizations
PDr Furkan, good day. Can you explain to me, in a dummy way, how can I edit the BAT file to have 2 folders with models in A1111 (original and another one) and 2 folders with LORAs (original and another one)?
Fprobably :d
Fcan you explain me more what you wan to do? i can show you
PI don't have a huge drive so I will divide my LORAs and my CHECKPOINTs in two diferent folders, one folder for LORAs and CHECKPOINTs in the drive D and other folder with other LORAs and CHECKPOINTs in the drive E. I want to have two separeted folders in two different drivers with the same installation of A1111. For this reason I want to edit the BAT file but I don't know how to.
Flet me show
Fchange your paths accordingly
Pand this way it will keep the original folder plus this one?
Fi think it only uses 1 folder
Fmultiple folders would be good i agree
PThanks
F@Dr. Furkan Gözükara Do i still need to downgrade to v21.8.4 of Kohya SS when training dreambooth on SD 1.5? Found from your patreon: https://www.patreon.com/posts/best-stable-1-5-90307694. I've been using your settings for SDXL dreambooth, but found that there did not exists any openpose for SDXL and i need to control the pictures generated.. Will try with SD 1.5
Patreon
Patreon is empowering a new generation of creators.
Support and engage with artists and creators as they live out their passions!
Support and engage with artists and creators as they live out their passions!
Fi think they fixed
Falso our latest SD 1.5 config works on latest version
Pi made a clean install of KOhya and now i dont get error when LORA extract. Git pulling oldd version didnt help. Only clean reinstall for ome reason
Fnice at least solved
FWas that the config in the Patreon post that I linked?
NModel called epiCPhotoGasm last version is pretty good
Bshould all stage1 references be remove from the GUI?
FWe may remove not useful
FPatreon
Patreon is empowering a new generation of creators.
Support and engage with artists and creators as they live out their passions!
Support and engage with artists and creators as they live out their passions!
FNice can be tested
A@Dr. Furkan Gözükara is there any way to rollback to supir v14 ?
Aits taking twice as long for me on the latest version and worse results
Rhi @Dr. Furkan Gözükara what's this gradio error?
Ralso no module called requests.
Ri installed your v19 and threw in the models
BI don't have v1.4 installed, but v1.7 quality is better than v1.9
Fthere shouldnt be this stuff
Fare you aware not it default selects Linear CFG?
Fi can turn it off
Flooks like same folder name repeat error
Finstall into different folde rname
Fcan you test turning off Linear CFG and compare?
Fthat could be reason
Fcan you test turning off Linear CFG and compare?
that could be reason
that could be reason
Fplease let me know i will change default according to that
Fcan you provide me your tested image and prompt? i can compare locally and see
Fwe have to fix if there is a quality drop
K@Dr. Furkan Gözükara Sora is using Diffusion transformers, but the use of those is nothing new. The quality of the output is something else though.
Pixart-alpha is also using diffusion transformers. Have you looked anything into this? https://github.com/PixArt-alpha/PixArt-alpha
Pixart-alpha is also using diffusion transformers. Have you looked anything into this? https://github.com/PixArt-alpha/PixArt-alpha
GitHub
Fast Training of Diffusion Transformer for Photorealistic Text-to-Image Synthesis - PixArt-alpha/PixArt-alpha
Fi know pixart alpha but didnt look carefully yet
Fi am adding dynamically shown target output resolution
