the full model uses like 28 gb vram
the full model uses like 28 gb vram
Fso both fits into vram
SThanks 
SIm using this preset configs: "Batch_Size_7_48GB_GPU" -- what GPU on runpod do you reccomend for this: L40s kinda sucks
[DreamBooth Flux Krea]
[DreamBooth Flux Krea]
FA100 faster than L40S
For H100
Fif you want better speed
K@Dr. Furkan Gözükara i was wondering if doing a finetune on krea is a bit different, the woman with teal braids is my finetuned model. Im doing depth to image. workflow i see a lot of realism and shadows on the krea model that gets lost when using my finetuned character krea.
Scan you share the workflow?
Sand i also understand that with classic finetune flux there is no issue?
Kthe depth one or the image to image?
Kim not sure if the classic finetune has that problem, but on krea is really obvious
C@Dr. Furkan Gözükara Dr. Furkan. Do you think that training on the flux krea model in Kohya GUI DreamBooth will be the same as in your tutorial from 22.10.2024:
https://youtu.be/FvpWy1x5etM?si=CI-dObKmDprZesoN
Should I just replace flux.dev with Krea?
Or maybe during this time, some significant optimizations have been introduced that would also be worth changing? Any other options?
Are you planning to do a comparison of flux.dev vs Krea in DreamBooth training?
https://youtu.be/FvpWy1x5etM?si=CI-dObKmDprZesoN
Should I just replace flux.dev with Krea?
Or maybe during this time, some significant optimizations have been introduced that would also be worth changing? Any other options?
Are you planning to do a comparison of flux.dev vs Krea in DreamBooth training?
YouTubeSECourses
If you want to train FLUX with maximum possible quality, this is the tutorial looking for. In this comprehensive tutorial, you will learn how to install Kohya GUI and use it to fully Fine-Tune / DreamBooth FLUX model. After that how to use SwarmUI to compare generated checkpoints / models and find the very best one to generate most amazing image...
JWan Training 
Dis it a bad idea when training a character lora (SDXL) to use two sets of the same images - one in 1024x1024 and one in 896x1152 to be more flexible depending on which aspect ration/resolution i want to create my images with? and if it is not a bad idea, should i create one concept per resolution? would i have to use 2 concepts with reg images?
Dand is there a benefit or downside in using full resolution images (in the correct aspect ratio) or should i downscale everything to the exact resolution (1024x1024/896x1152) or does it downscale on its own when the source images are higher resolution
Fi will test hopefully
Fyes i will do detailed test
Flots of shared vrma
Fit must be very slow
Fostris still doesnt support block swap
Finteresting did he add block swap?
Fit is not as slow as much i would say
Fwhat is your training resolution?
Fi would say make all images exact resolution
Fyou probably can train with 2 different aspect ratio
Fmake sure to enable bucketing
Dthank you. would i need both sets of reg images? so in the end 2 image concepts and 2 reg image concepts? or would you just throw them all in one folder (1 image folder, 1 reg folder)
Fgood question. probably having different aspect ratio would help
Fyou training sdxl right?
Dyes sdxl
Fyes it should work
Dcurrently i have two folders, one with 1024x1024 and one with 896x1152. so the question is: should i create one concept in onetrainer with both resolutions in one folder, or separate them and have two concepts
Dand would i add 2 reg image concepts for either resolution
Dor one concept with both reg resolutions inside
Fin one trainer
Fjust add them as 2 concepts
Fand properly set their repeat count
Dso 2 concepts for the images (1024x1024 and 896x1152). i have 50 images each. since it's 2 concepts do i use 50 or 100?
Flogic simple
Fat 1 epoch you have 100 train images
Fthen you should use 100 reg images
F50 from each in that case
Jdue to block swap and vae_cache_cpu enabled.
interestingly enough it made no difference in my case for speed whether i turn it on or not for that 480z360 dataset , tested speeds with and without. (without it fits in dedicated vram without shared mem as well)
(though usualy vram bottleneck does limit speeds in other instances)
interestingly enough it made no difference in my case for speed whether i turn it on or not for that 480z360 dataset , tested speeds with and without. (without it fits in dedicated vram without shared mem as well)
(though usualy vram bottleneck does limit speeds in other instances)
Ji am still messing around with what barely fits in vram, apparently as long as there is about 1gb vram dedicated available it will run full speed (400 watt at 70% power limit)
JFor Wan2.1 T2V using (25 videos) 480x360@61 and (23) images 960x720 fits just perfectly in an RTX 5090. Wondering about the results, but this is the most I was sable to cram into the VRAM without going overboard and not using any offloading / shared vram
J
CYesterday, I extracted a LORA from my Dreambooth model (of my character) that I trained in Kohya GUI on the FLUX.dev model. I wanted to see how the LORA would perform with the Flux Krea model, but unfortunately, it doesn't look great.
I guess I'll have to train a new LORA specifically on the Flux Krea model, or train an entire Dreambooth model from scratch.
I guess I'll have to train a new LORA specifically on the Flux Krea model, or train an entire Dreambooth model from scratch.
JMakes sense I guess
AI can confirm. Lora trained on Dev inferenced on krea is not that good
ABut a checkpoint with krea Lora works
