Software Engineering Courses (SECourses)•11mo ago

@Furkan Gözükara SECourses what would be the way to kind of automate the flux finetuning process mor

@Furkan Gözükara SECourses what would be the way to kind of automate the flux finetuning process more? is there a way to create an image that has fixed flux version etc that you can just load up on cloud computer?
this would not be possible with massed compute?
with runpod?
or is there a way to make this more automated.... not exactly sure how... but maybe with lambda from AWS? any idea?

VVioletAnt @Furkan Gözükara SECourses what would be the way to kind of automate the flux fi...

Furkan Gözükara SECourses•2/20/25, 9:55 PM

automate = software development skills or $$$$$

FFurkan Gözükara SECourses automate = software development skills or $$$$$

VioletAntOP•2/20/25, 10:06 PM

haha... yeah I got a computer science degree and cloud certification

but not sure yet which are the plattforms or tools to use.... so can tell me which are the suited plattforms for quick protoyping, also cheap would be good.... because it seems massed compute is cheapest but has super limited resources and you cannot upload prebaked images? Didnt have a look at how runpod works yet... Before somebody mentioned somehting about "replicate"....

I mean I could create an image for AWS instance.... then add api gateway, lambdas etc... but I have to look actually how much these gpus cost on AWS platform directly... will have a look tomorrow regarding pricing...
but then I would need to create some kind of application with api running on the aws instance which then creates the images etc...

VVioletAnt haha... yeah I got a computer science degree and cloud certification 😄 but not ...

Furkan Gözükara SECourses•2/20/25, 10:22 PM

i never researched to be frank

FFurkan Gözükara SECourses i never researched to be frank 😄

VioletAntOP•2/20/25, 10:23 PM

hmmm... yeah maybe the replicate api... then combine it with some other things... I guess there are also some turnkeysolutions for user interfaces.... Need to look a bit more

from replicate:
Nvidia L40S GPU (gpu-l40s, $3.51/hr)
Nvidia A100 GPU (gpu-a100-large, $5.04/hr)
but it says the api use is billed per second

VioletAntOP•2/20/25, 10:28 PM

Actually somebody saw the pictures I generated with the flux and wants me now to create some images for him for money so yeah....

rob•2/20/25, 10:28 PM

Hey Doc, I just saw the IDM VTON app. Have you compared IDM VTON with CAT VTON and/or Flux Redux?

VVioletAnt hmmm... yeah maybe the replicate api... then combine it with some other things.....

Furkan Gözükara SECourses•2/20/25, 10:30 PM

ye people using replicate

Rrob Hey Doc, I just saw the IDM VTON app. Have you compared IDM VTON with CAT VTON a...

Furkan Gözükara SECourses•2/20/25, 10:30 PM

i havent compared with cat vton yet

Furkan Gözükara SECourses•2/20/25, 10:30 PM

flux redux works different

FFurkan Gözükara SECourses The grid test I am analyzing to find newer FLUX DEV training. This is not even i...

Neuromasters•2/21/25, 7:56 AM

You want to share?
I'll be sure to test it
And here's a question, has Hunyuan3D-2Hunyuan3D-2 been deployed yet? Who knows can advise solutions for 4090rtx?
https://huggingface.co/spaces/tencent/Hunyuan3D-2/tree/main

FFurkan Gözükara SECourses ye i read it

�

𝗡𝗶𝘁𝗶𝗻 𝗣𝗮𝗿𝗲𝘃𝗮•2/21/25, 8:04 AM

Kohya implemented this - https://github.com/kohya-ss/sd-scripts/pull/1864 can you please tell me how to start this in kohya. It will be a great factor to analyze how the lora or dreambooth is training.

GitHub

Validation loss by rockerBOO · Pull Request #1864 · kohya-ss/sd-sc...

Related #1856 #1858 #1165 #914
Original implementation by @rockerBOO
Timestep validation implementation by @gesen2egee
Updated implementation for sd3/flux by @hinablue
I went through and tried to m...

julius•2/21/25, 9:55 AM

trying to generate 4 images at the same time (using 4 L40) but it seems like it will only make 2 at a time?

If I leave OverQueue=1 (default) then 2 will generate exactly the same time, and then the next 2.

If I set OverQueue=0 then it still takes about 40 seconds before the fourth image starts to generate, so ends up taking just as much time. Like the first photo will be finished before the fourth one starts.

Shouldn't all 4 photos just start generating at the same time?

If I ask to generate 8 images at the same time, all 4 GPU's will run from the start like it should though...

On Massedcompute.

Jjulius trying to generate 4 images at the same time (using 4 L40) but it seems like it ...

Timson•2/21/25, 12:43 PM

try looking at debug logs

TTimson try looking at debug logs

julius•2/21/25, 1:52 PM

ah yea should've sent that. forgot to look and have closed down now

Pelolisu•2/21/25, 2:31 PM

Hi guys, I'm testing a new upscaler technique and I'd like your opinion on the result, what do you think? I'm going to send 2 images, one in 1024x768 and the other in 4096x3072 upscaler.

Pelolisu•2/21/25, 2:31 PM

Pelolisu•2/21/25, 2:32 PM

this 4k https://drive.google.com/file/d/1KdOPSf7uPagU9zTgrHSLWy8OKSorKeAV/view?usp=sharing

Google Docs

mx_upscaled_image.png

PPelolisu this 4k https://drive.google.com/file/d/1KdOPSf7uPagU9zTgrHSLWy8OKSorKeAV/view?u...

Timson•2/21/25, 3:01 PM

looks decent

Timson•2/21/25, 3:12 PM

Does anyone use KLING AI paid memberships?
I was looking to get yearly sub, but they change the deal dramatically just over the course of a day removing daily log in credits (which is basically x2 potential monthly generations)
The reviews on reddit also indicate that the support is non existent and many people face problems with one sided deal changing and zero communication with clients

Timson•2/21/25, 3:32 PM

As for quality Kling AI 1.6 seems to be the best img2vid model publicly available as of now

Timson•2/21/25, 3:34 PM

Closest competitors hailuoai and sora currently seriously behind

TTimson As for quality Kling AI 1.6 seems to be the best img2vid model publicly availabl...

AiInfluence•2/21/25, 4:30 PM

it's the best currently what do you mean they actually did remove ??

AiInfluence•2/21/25, 4:31 PM

I suppose their abusing monopoly until a competitor who comes beat them

FFurkan Gözükara SECourses 8GB_GPU_Quality_Tier_4_FP8_6850MB_13.2_Second_IT.json

Guy Bertental•2/21/25, 4:40 PM

Hi @Furkan Gözükara SECourses,
I've tried both flux1-dev-fp8.safetensorsflux1-dev-fp8.safetensors and flux1-dev.safetensorsflux1-dev.safetensors based on 8GB_GPU_Quality_Tier_4_FP8_6850MB_13.2_Second_IT.json8GB_GPU_Quality_Tier_4_FP8_6850MB_13.2_Second_IT.json config
on 48 images with faces pre-captioned with joy captioning and replaced with ohwxohwx token,
it takes too much time - estimated 45 days! is that make sense??
Shared GPU raised to 14.2 and GPU RAM is on 7.5 GB.
I've attached all screenshots

GGuy Bertental Hi @Furkan Gözükara SECourses, I've tried both `flux1-dev-fp8.safetensors` and `...

Guy Bertental•2/21/25, 4:42 PM

another screenshot with more lines from training console

NNeuromasters You want to share? I'll be sure to test it And here's a question, has `Hunyuan3D...

Furkan Gözükara SECourses•2/21/25, 5:48 PM

i will share hopefully after finalizing

V3ldrin•2/21/25, 5:49 PM

Hi, I also have a problem when training with kohya: I have a geforce rtx 3900 with 24 vram, I use the models from your youtube tutorial video (flux dev, fp16, etc.etc). Before training, I clear vram (restart, autostart apps off, etc). When I check, I have enough vram under 500 mb. When I start training, it goes up to 23 GB and ALSO the shared Vram goes up. Why? Is that normal? using the 24GB_GPU_Quality_Tier_3_23150MB_5.9_Second_IT