FLUX LoRA: Free Hollywood-Quality Art at Home on Your PC - Unleash Your Imagination & Become a Hero

[@SECourses]

These images are generated with FLUX LoRA training for free. Full workflow is as follows: 🔗 Medium article ⤵ ▶ medium.com/@furkangozukara/ultimate-flux-lora-training-tutorial-windows-and-cloud-deployment-abb72f21cbf8 🔗 Windows main tutorial ⤵ ▶ kzbin.info/www/bejne/pKq2eKhnZ75ja80 🔗Cloud tutorial for GPU poor or scaling ⤵ ▶ kzbin.info/www/bejne/Y6bLfWWkjJx3mtk 🔗 Full detailed results and conclusions ⤵ ▶ www.patreon.com/posts/111891669 🔗Full config files and details to train ⤵ ▶ www.patreon.com/posts/110879657 🔗SUPIR Upscaling Tutorial ⤵ ▶ kzbin.info/www/bejne/hYrbh3isectlbpY

[@LouisGedo]

👋 hi

[@SECourses]

@@LouisGedo hi

[@Abhijeetshalgar]

​@@SECourses how much speed of internet required for rent gpu in cloud

[@Abhijeetshalgar]

@@SECourses how much speed of internet required for rent gpu in cloud

[@SECourses]

@@Abhijeetshalgar very minimal. you should be able to use both runpod and massed compute very well

[@Beauty.and.FashionPhotographer]

the Flux plastic face skin problem can be fixed if you use "iphone picture"... or you use "medium format shot", "Shot with hasselblad camera" , "shot on kodak Ektachrome 100 film", use all or a mix of them....

[@SECourses]

ye that is a good solution. also sufficient training removes that plastic skin problem as well

[@Alphfirm]

Wow first time time AI images really are on point! So cool you've stuck with this work for so long!

[@SECourses]

thanks a lot. even better ones are on the way with perfected fine tuning hopefully :)

[@Alphfirm]

@@SECourses Sounds awesome! Especially if cloud/runpod guides :)

[@SECourses]

@@Alphfirm you are welcome. ye you can run it same on runpod massed compute. i think i can show there too quickly

[@CaraDePatoGameplays]

now I can be a financial influencer who lives around the world with a notebook without leaving home

[@Dany-w3g]

You are a real Monegasque!

[@SECourses]

haha true :D

[@UnchartedWorlds]

100% 😂😂😂

[@Abhijeetshalgar]

​@@SECourseswhich GPU have 10k resolution

[@SECourses]

@@Abhijeetshalgar get rtx 4090

[@118khw]

Hi, thanks for the great content! If I follow your instructions to use SwarmUI in Runpod, should I stick to Pytorch 2.1 template? OR should I use 2.4? Thank you

[@SECourses]

stick pytorch 2.1 it works great

[@iseahosbourne9064]

Hey SECourses, Im trying to make animations right now. Say, if I create the scene in a 3D software like blender, then render it out to images, what would be the best workflow to stylize / cartoonise the enviroment and characters present, seperately, with depth and temporal consistency, aka as realistic to animation as possible? Viggle, animate diff or something else?

[@SECourses]

for animating not much thing. but if you want to use image to image SDXL is better with controlnets

[@iseahosbourne9064]

@@SECourses Comfy UI?

[@SECourses]

@@iseahosbourne9064 i am not really expert at comfyui i use swarmui most :D

[@frustasistumbleguys4900]

Can you create a workflow for two people using different LoRa, such as a man and a woman? or multi people there? just in case.

[@SECourses]

currently that doesnt work well. you need to have pictures of 2 person in the same image. that time it works. i may make a tutorial for this

[@UnchartedWorlds]

It's easy with photoshop workflow and inpainting lora, I've done this with photorealistic amazing results. Two people, two different loras and you do multiple rounds of inpainting and overlays in photoshop until you have satisfying result.

[@SECourses]

@@UnchartedWorlds yes you can do this as well. good tips

[@UnchartedWorlds]

Doctor Ai, I mean professor I've noticed that with flux and loras the animals like dinosaurs or dragons can look super plastic, and you have similar result at 0:21,any ideas why? And what could be done about for more realistic skins?

[@SECourses]

you are 100% right. i couldn't find a solution yet :/ maybe a fine tuned FLUX is necessary

[@UnchartedWorlds]

​@@SECourses​looking forward you training realistic dragon and dinosaur lora and releasing it to the world! :)

[@SECourses]

@@UnchartedWorlds ye that would be nice

[@sphiweapril316]

How are you always getting clear backgrounds? My generations always come out with shallow depth of field which is soft and often blurry

[@SECourses]

the LoRA training making that impact. As you do train more epochs, that annoying "shallow depth of field which is soft and often blurry" effect is gone. But more training also causes overfitting so you have to have a very good config + decide which checkpoint to use for balance both

[@tazztone]

@@SECourses what settings did you change to train with your new dataset? or the same as "Rank_3_18950MB_9_05_Second_IT"?

[@SECourses]

@@tazztone ​@tazztone it was rank 1 with T5 XXL on 8x GPU so the LR was double of rank 3, not other diff

[@tazztone]

@@SECourses thanks. i thought only the UNET was trainable 🤔? so text encoder can be trained too? so you end up with a fine tuned t5 xxl safetensor file?

[@SECourses]

@@tazztone ​@tazztone yes I trained both T5 XXL and also clip l, they also get added to the Lora file, only related weights not fully

[@RANJEET3939]

Hi, why didn't you use comfy UI and AI toolkit for training and generating images?

[@SECourses]

Because Kohya and SwarmUI are easier and working perfect

[@fluffsquirrel]

This is so cool, fantastic demonstration! For some of these, I could easily confuse them for real photographs. Doesn't seem to be many artifacts except for the eye of the black and white tiger.

[@SECourses]

good catch there :) and yes this is super high quality

[@fluffsquirrel]

@@SECourses Thank you!

[@SECourses]

@@fluffsquirrel you are welcome and thanks for comment

[@MikesSoundStudio]

Great job :), Where can i found these prompts? I am a member of Patreon :)

[@SECourses]

Download latest zip file and inside Test_Prompts file name is 28 Prompts Yolo Face.txt

[@alex_great23]

please make a scheme comfyui for Flux where you can load 2 loras of different people and put them into one category.

[@SECourses]

i, use SwarmUI and it works amazing

[@solomslls]

Good video, can you please update your config for kohya ss for flux to include ones for 512x512 images, I'm using your Patreon ones, but my GPU usage is at 7gb when training with your 12gb script , feel like my GPU can handle more , it never goes up to even 10

[@SECourses]

your strategy should be like this. start from like rank 9 and raise your rank until your gpu starts using shared vram. 512 reduces vram usage a lot so it is a special case. so you should try ranks with 512px. lower rank better quality rank 1 better than rank 2 and so on

[@solomslls]

@@SECourses thx I'll try it , and hope you include 512 and 768 scripts next time, almost all flux Lora uploaded on civitai are trained on 768 , people don't train on 1024 , thx again 🙏

[@SECourses]

@@solomslls but 1024 yields best quality , if you aim quality you shouldnt reduce :D 512 works good too of course but not as good as 512, i will hopefully add 512 px configs

[@solomslls]

@@SECourses yeah I agree, most flux users now are old sd1.5 and sdxl users, 512 training can be done on 8gb too, and thx man appreciated 👍,

[@neogeo9117]

@@solomslls We've been training on 1024px since SDXL...and some before that on custom SD 1.5 and 2.0 base models trained on 1024px datasets...

[@mr.entezaee]

Why, for example, the style of 3d , line painting or watercolor or other styles can't be seen in the photos, and they are all real? Or if there is, it seems that he has changed his face very little ... Should there be an artistic style in the photos we show first? Or it doesn't matter

[@SECourses]

currently it is either fully stylized or realistic. FLUX can't do both atm :/ so for stylized output you need to train lesser epoch, but for realistic you have to train more epoch

[@brandonalan]

*¡BRAVO!* 💯👏

[@SECourses]

thank you so much

[@LinusBuyerTips]

it's just beginning

[@SECourses]

Yep. it is getting better and better

[@Mosen_xd]

perfect x2👌

[@SECourses]

thank you so much

[@abhinavbisht9851]

Amazing doc 😻

[@SECourses]

thank you so much

[@smartkhawar]

great lecture.

[@SECourses]

thanks a lot

[@dejanp8558]

🔥🔥🔥

[@SECourses]

Thanks for the comment

[@1Know1tHurts]

I will be testing new configs tomorrow, thanks!

[@SECourses]

great let me know results

[@gumvue.studio]

perfect.. now animate

[@SECourses]

for animating sadly there isnt any good open source model. only closed paid ones

[@iddeegee]

ha ha, nice :)

[@SECourses]

thanks a lot :D

[@jonotonfoto]

While the author previously had lengthy videos, they contained interesting ideas and insights, for which I'm grateful. Now, however, you'll waste a lot of time and won't find answers because they're now behind a paywall. All worthwhile content is now on Patreon. This long and drawn-out video is just an advertisement for their Patreon. Unfortunately, the author hasn't found a balance between public interest (which they've benefited from) and their own profit

[@SECourses]

I still give tons of information in the videos, if you think that 90 minutes of tutorial is entirely paywalled you are very wrong.