Software Engineering Courses (SECourses)•9mo ago

number of samples meaning your training dataset size?

AikageOP•4/4/25, 1:51 PM

or how many times it should read

AikageOP•4/4/25, 1:51 PM

i am bad with terminology

Inoshi•4/4/25, 4:02 PM

can i use flux lora with the flux fill model ?

FFurkan Gözükara SECourses you can use swarmui auto downloader

xlads•4/4/25, 5:45 PM

so i want to download flux.dev to use for general image generation which is the link to download? i already have swarm ui installed in my pof

xlads•4/4/25, 5:51 PM

so i jsut pasted this ""C:\swarm_ui\SwarmUI\Models\Stable-Diffusion\OfficialStableDiffusion\sd_xl_base_1.0.safetensors" file which i had inside of my swarmui that is installed locally into my pod hopefully that will work. but i dont see anything insdie of the flux file inside of my local swarm ui download for me to paste it into my pod model / stable-diffusion file

xlads•4/4/25, 6:01 PM

okay so ive downloaded the flux.dev model but the images are coming out sort of blurry any way to fix this ?

1800001-A_group_of_friends_sprawled_on_a_blanket-flux1-dev-fp8.png

Kompicka•4/4/25, 6:37 PM

Any tutorial for using yoloface? All results with it just result in grain in total image. I have correct folder and correct download.
EDIT:Fixed it using the new womanfacev5_1024.pt . Still dont know why face_yolov9c didnt work.

JaviEM•4/4/25, 8:39 PM

@Dr. Furkan Gözükara Is there a way to use gguf models in your version of wan 2.1? Downloading the models manually or something. In the ComfyUI version everyone can do it from almost day one.

AAikage number of samples meaning your training dataset size?

Furkan Gözükara SECourses•4/4/25, 9:06 PM

number of samples means how many samples model see during training

Furkan Gözükara SECourses•4/4/25, 9:06 PM

like 10 repeat 5 epoch 3 image = 10 * 5 * 3 = 150 samples

Xxlads okay so ive downloaded the flux.dev model but the images are coming out sort of ...

Furkan Gözükara SECourses•4/4/25, 9:07 PM

set cfg 1

KKompicka Any tutorial for using yoloface? All results with it just result in grain in tot...

Furkan Gözükara SECourses•4/4/25, 9:07 PM

weird

JJaviEM @Dr. Furkan Gözükara Is there a way to use gguf models in your version of wan 2....

Furkan Gözükara SECourses•4/4/25, 9:07 PM

not yet but i will add soon hopefully. it just arrived to the library we use

Kompicka•4/4/25, 11:28 PM

Should i change any particural settings in SWARMUI when using the FP8 converted dreambooth checkpoint? I have considerable drop in quality and it should not be recognizable.

KKompicka Should i change any particural settings in SWARMUI when using the FP8 converted ...

Furkan Gözükara SECourses•4/4/25, 11:30 PM

you dont need to change

Furkan Gözükara SECourses•4/4/25, 11:30 PM

remember guys when using swarmui segmentation you can change sort order

Furkan Gözükara SECourses•4/4/25, 11:30 PM

like yolo face etc

FFurkan Gözükara SECourses no it doesnt need

sergio_gonzo•4/5/25, 12:14 AM

dr Furkan for 66 image how many epoch do you tink could be enoug?

sergio_gonzo•4/5/25, 12:14 AM

train batch size of 1

Ssergio_gonzo dr Furkan for 66 image how many epoch do you tink could be enoug?

Furkan Gözükara SECourses•4/5/25, 12:35 AM

i would train up to 150 and compare

Furkan Gözükara SECourses•4/5/25, 12:39 AM

i got to sleep now.

sergio_gonzo•4/5/25, 12:39 AM

thank you dr furkan, i put train bath in 2 also

Ssergio_gonzo train batch size of 1

Kompicka•4/5/25, 12:59 AM

Hi just for your information. I have pretty curated dataset with 97 images of character. Best epochs are between 55 and 70. Still testing and not sure.
So 97 * 55 around 5500 - 7000 steps using the dreambooth config.

Kompicka•4/5/25, 1:02 AM

But everyone has different dataset so train what you can, just the testing a lot of checkpoints is very tedious since they behave differently based on the promp.

Kompicka•4/5/25, 1:14 AM

I think i messed up using my dataset with captions. Now i need to use my captions to generate great pics, but that diminishes flexibility.

Kompicka•4/5/25, 1:17 AM

All in all still up one level from over 200 loras I have trained. Dont know why i have never tried dreambooth training

sergio_gonzo•4/5/25, 2:26 AM

THAnk you for your insight i will test in another what you say

Kompicka•4/5/25, 2:29 AM

Good luck, what are you training?

sergio_gonzo•4/5/25, 2:38 AM

a actress

Kompicka•4/5/25, 2:55 AM

i think batch size 1 is always best for likeness, at least based on my testing.

sergio_gonzo•4/5/25, 2:58 AM

but for a better understanding

sergio_gonzo•4/5/25, 2:58 AM

2 is the ha;ft time that batch 1?

�

𝗡𝗶𝘁𝗶𝗻 𝗣𝗮𝗿𝗲𝘃𝗮•4/5/25, 9:20 AM

@Furkan Gözükara SECourses I am training Wan 2.1 with sage attention and it's quite fast. Can we train flux with sage attention?

Aragon•4/5/25, 11:51 AM

If I have a Flux.1-dev character finetune (myself) checkpoint, and I'd like to start using different LoRAs, is it better to use these LoRAs on my fintune checkpoint, or would I get better results if I extract my checkpoint into a LoRA, and then use the original Flux model with both my character LoRA and the other LoRAs?

The reason I'm asking is that I'm thinking that maybe since the LoRAs I'll use were trained on the original model, maybe using them on a finetuned model could cause unexpected / less accurate results?

Original message was deleted

Aragon•4/5/25, 1:02 PM

@Furkan Gözükara SECourses please ban this dude and his previous alt account he used

AAragon @Furkan Gözükara SECourses please ban this dude and his previous alt account he ...

Furkan Gözükara SECourses•4/5/25, 1:02 PM

ye banned

Furkan Gözükara SECourses•4/5/25, 1:02 PM

looks like hacked

�𝗡𝗶𝘁𝗶𝗻 𝗣𝗮𝗿𝗲𝘃𝗮 @Furkan Gözükara SECourses I am training Wan 2.1 with sage attention and it's qu...

Furkan Gözükara SECourses•4/5/25, 1:03 PM

i think not

Furkan Gözükara SECourses•4/5/25, 1:04 PM

it has a function though : https://github.com/kohya-ss/sd-scripts/blob/a0f11730f744ce48d3a7a9005b783317b753cc0c/library/attention_processors.py#L15

GitHub

sd-scripts/library/attention_processors.py at a0f11730f744ce48d3a7a...

Contribute to kohya-ss/sd-scripts development by creating an account on GitHub.

Furkan Gözükara SECourses•4/5/25, 1:04 PM

for flash attention

AAragon If I have a Flux.1-dev character finetune (myself) checkpoint, and I'd like to s...

Furkan Gözükara SECourses•4/5/25, 1:05 PM

your logic is valid

Furkan Gözükara SECourses•4/5/25, 1:05 PM

loras works best on trained model

Furkan Gözükara SECourses•4/5/25, 1:06 PM

i couldnt do much this week, it was holy day and also i am trying to finish this app

404_error•4/5/25, 4:59 PM

I want to fine tune a Flux model with my face, i see on this video, reg image do not improve quality, is that true ?
https://www.youtube.com/watch?v=nySGu12Y05k&list=WL&index=60&ab_channel=SECourses

YouTubeSECourses

FLUX LoRA Training Simplified: From Zero to Hero with Kohya SS GUI ...

Ultimate Kohya GUI FLUX LoRA training tutorial. This tutorial is product of non-stop 9 days research and training. I have trained over 73 FLUX LoRA models and analyzed all to prepare this tutorial video. The research still going on and hopefully the results will be significantly improved and latest configs and findings will be shared. Please wat...

1nfern0•4/5/25, 6:29 PM

what I wouldn't do for one of these beasts https://www.scan.co.uk/products/96gb-pny-nvidia-rtx-pro-6000-blackwell-24064-cuda-752-tensor-188-rt-gddr7-w-ecc-pcie-50x16-4x-dp-21

PNY NVIDIA RTX PRO 6000 96GB Blackwell Workstation Edition Retail G...

11nfern0 what I wouldn't do for one of these beasts https://www.scan.co.uk/products/96gb-...

.•4/5/25, 7:20 PM

4404_error I want to fine tune a Flux model with my face, i see on this video, reg image d...

Kompicka•4/5/25, 7:36 PM

hi i havent seen any training with reg images for flux, they were used in SD. I also think Furkan does not use them for Flux as well. At least I havent seen it used anywhere.

Kompicka•4/5/25, 7:37 PM

I also dont think you will overfit dreambooth easily, so it seems redundant

404_error•4/5/25, 7:37 PM

Thanks you

4404_error I want to fine tune a Flux model with my face, i see on this video, reg image d...

Furkan Gözükara SECourses•4/5/25, 10:29 PM

yes for flux multi subject gets mixed

number of samples meaning your training dataset size?

Similar Threads

Similar Threads

Similar Threads