Stable Diffusion 3.5 Large How To Use Tutorial With Best Configuration and Comparison With FLUX DEV

[@SECourses]

🔗 Full Open Access Public Article Used In The Video ⤵ ▶ www.patreon.com/posts/106135985 🔗 Unified Models 1-Click Robust Auto Downloader Scripts ⤵ ▶ www.patreon.com/posts/114517862 🔗 Full FLUX Tutorial - FLUX Beats Midjourney for Real ⤵ ▶ kzbin.info/www/bejne/mKbTg5iGirR0Z5o 🔗 Main Windows SwarmUI Tutorial (Watch To Learn How to Use) ⤵ ▶ kzbin.info/www/bejne/fny7aZJ8Zqqlldk 🔗 Cloud SwarmUI Tutorial (Massed Compute - RunPod - Kaggle) ⤵ ▶ kzbin.info/www/bejne/jne4i6Kca7ieodk 🔗 SECourses Discord Channel to Get Full Support ⤵ ▶ discord.com/servers/software-engineering-courses-secourses-772774097734074388 🔗 SECourses Reddit ⤵ ▶ www.reddit.com/r/SECourses/ 🔗 SECourses GitHub ⤵ ▶ github.com/FurkanGozukara/Stable-Diffusion

[@BrahmandaBytes]

Great. Do we need to follow the steps every time we boot in runpod? or the installed models persist?

[@SECourses]

if you stop pod and not terminate it will stay but will use your credits. i recommend this tutorial : kzbin.info/www/bejne/boLJZatpYql5nJY

[@firatguven6592]

Furkan kardeşimin en sevdiği kelimeler it's Amazing, it's the best 😉

[@SECourses]

:D

[@admiralevan]

hey bud, Im using the medium SD3.5 on SwarmUI with your settings, but it looks terrible, do you know why?

[@SECourses]

the model is not very good. so you get worse results than me with same prompts?

[@NamKim-b1c]

which service do you recommend, mimicpc or Massed Compute?

[@SECourses]

I recommend Massed Compute. MimicPC is easy to use but expensive.

[@johnnyapbeats]

Hello Dr. Furkan great video and value as always, i would love to see a tutorial on how to ultimately prepare and train Product Lora's for a small / medium business!

[@SECourses]

that is on my list hopefully. as a pre requirement make sure to watch this video : kzbin.info/www/bejne/fKfTiKxnrZqYqq8

[@johnnyapbeats]

@@SECourses you are the man Dr., thanks for your valuable contribution 🙏🏼

[@SECourses]

thank you so much

[@sinayagubi8805]

on Massed Compute, can we pay for permanent storage so we don't have to download the loras and models again and again?

[@SECourses]

sadly not possible yet. i will forward this right now

[@UnlimitedGRenemy]

do you work specifically with SWARM ui or do you also have workflows for comfy let say

[@SECourses]

In past I was most focused on Automatic1111 but after FLUX mostly SwarmUI. I don't focus on ComfyUI sadly

[@SECourses]

@@user-yo8pw8wd3z sure it is i agree. but it is just harder to use. i try to balance between very hard and medium - by the way since swarmui uses comfyui as back end, it gets all same updates

[@UnlimitedGRenemy]

@@SECourses Another question because i only started using stable diffusion recently, how do you train such model locally, do you have a link or will you prepare a detailed guide?

[@SECourses]

@@UnlimitedGRenemy yep here full local training guide : kzbin.info/www/bejne/fKfTiKxnrZqYqq8

[@UnlimitedGRenemy]

@@SECourses so is the same way as Kohya ?

[@digidope]

Surprinsingly SD 3.5 is un-censored out of the box. But why compare it to dev and not pro 1.1?

[@SECourses]

1.1 is paid and extremely limited, with dev you can fine tune train lora do whatever you want

[@digidope]

@@SECourses For non GPU owners all are paid. Out of the box 1.1 outperforms SD 3.5 in all areas except on censoring. In replicate 3.5 is $0.065/image and 1.1 is 0.05/image so its cheaper. DEV is wildly different than 1.1. Is that good or bad depends what you do. Replicates new playground is the most easiest place to test every model at the same time. It now has both Ideogram models too.

[@SECourses]

@@digidope for non GPU owners i really recommend Massed Compute. 31 cents per hour and you run models as they are in your computer. but comparison of all models are good

[@sarpsomer]

Get well soon, another great tutorial!

[@SECourses]

thank you. yes i got some sickness

[@Varibam]

How do you make the SD 3.5 Images look so bad? I got a completely different experience. For Hands Flux > SD 3.5 no question. But in Terms of Photo realism, creativity, speed and color depth flux can go home.

[@SECourses]

well you see how i compared i didnt anything particularly :D and no SD 3.5 is behind

[@richardrispoli4508]

Very interesting comparison. Thx. 3.5 is still young but the differences are too big now. Flux dev has a strict licensing though compare to schnell or 3.5. Im more interested in schnell as i want to be able to use it professionally.

[@SECourses]

There is schnell de-distillation attempts. If they mature it may become commercially viable and working good.

[@richardrispoli4508]

@SECourses yes I'm following up this closely. You re probably one of the most advanced in terms of comparing all this models. I'm using AI to speed up game creation and keeping up with the pace of novelties is not easy.

[@SECourses]

@@richardrispoli4508 thanks. yes they are also in my radar will hopefully test each one

[@richardrispoli4508]

Distilled versions allows for fine tuning, that's what I understood from a previous video :) now everything makes sense 😅

[@SECourses]

@@richardrispoli4508 de-distilled models supposedly to allow multi-concept training we will see :D

[@evetevet7874]

Ccan u give me the name of your gpu model?

[@SECourses]

I got RTX 3090 TI

[@RANJEET3939]

mr. champion just posted :)

[@SECourses]

thanks a lot

[@umitbilen4007]

Bende diyorum bu kanaldaki ingilizceyi anlamam nasıl bu kadar kolay olabiliyor diye, meğersem hoca Türkmüş 😂

[@1Know1tHurts]

Only native speakers have hard time. I am a foreigner too and understand everything perfectly.

[@umitbilen4007]

I agree with what you say, it is very difficult to understand native English speakers because of their accent. For example, there is a speech of Adele on KZbin, I almost don't understand any word 😂​@@1Know1tHurts

[@umitbilen4007]

@@1Know1tHurts I agree with what you say, it is very difficult to understand native English speakers because of their accent. For example, there is a speech of Adele on KZbin, I almost don't understand any word 😂

[@SECourses]

@@1Know1tHurts true it is easier for foreigners to understand foreigners speaking :D

[@SECourses]

Aynen öyle :D

[@emircanerkul]

Türksün herhalde :)

[@SECourses]

Evet :)

[@evetevet7874]

hocam türkmüsünüz

[@SECourses]

Evet : www.linkedin.com/in/furkangozukara/

[@sternkrieger1950]

Photorealism on base Flux is BAD. Everything has this waxy, plasticy fake look to it with overblown contrast. You can especially see this with the dinosaur but humans are affected too. Finetuning and LORA fixes this, but Flux still isn't great because it is a distilled model with catastrophic forgetting. Have you tried training on SD 3.5 yet? I bet SD 3.5 may be actually beat Flux after training!

[@SECourses]

true. sd 3.5 research starting very soon hopefully

[@sdfwassdw9267]

I just trained a sd3.5lora ,rank 32 ,45MB ,it's really bad,

[@SECourses]

sad. hopefully it is my new research to compare. by the way did you see this? it is the absolute quality so far : kzbin.info/www/bejne/fKfTiKxnrZqYqq8

[@LarsBjurman]

Interesting! On the 3.5 Large model? How did you manage to do it?

[@SECourses]

@@LarsBjurman Kohya adding support and fixing bugs. almost ready.

[@LarsBjurman]

@@SECourses Ye, I know, also onto that, not working properly though, my training of 3.5 freezes atm. Was more interested to know how @sdfwassdw9267 made it work for him/her. Any clue would be much appreciated.

[@LarsBjurman]

​@@SECourses Yes, I know, not working properly for me yet though, Was curious if @sdfwassdw9267 had made it work. Any clue appreciated.