Software Engineering Courses (SECourses)•6mo ago

yeah, when I tried chroma the images weren't as good as flux dev, but on the bright side chroma has

yeah, when I tried chroma the images weren't as good as flux dev, but on the bright side chroma has apache license and I was wondering if finetuned well, will the results be in par with flux dev.

Uunderdog yeah, when I tried chroma the images weren't as good as flux dev, but on the bri...

Furkan Gözükara SECourses•7/21/25, 12:42 PM

kohya adding it

Furkan Gözükara SECourses•7/21/25, 12:42 PM

https://github.com/kohya-ss/sd-scripts/pull/2157

GitHub

Feature Chroma support by kohya-ss · Pull Request #2157 · kohya-s...

Thank you lodestones for the high quality model.
--model_type option is added to flux_train_network.py. For Chroma, --model_type chroma is required.
According to lodestones, --apply_t5_attn_mask --...

Le_Docteur•7/21/25, 4:20 PM

Mister Doctor, if i am to train on another SDXL model with baked vae, should i delete additional vae from preset? Tier1_48_GB_Faster_v2

this one
VAE (Optional: Path to checkpoint of vae for training)
stabilityai/sdxl-vae

LLe_Docteur Mister Doctor, if i am to train on another SDXL model with baked vae, should i d...

Furkan Gözükara SECourses•7/21/25, 4:43 PM

yes delete

FFurkan Gözükara SECourses yes delete

Le_Docteur•7/21/25, 4:58 PM

thank you
and if i do batch size=2, do i need to adjust learning rates? like double them

LLe_Docteur thank you and if i do batch size=2, do i need to adjust learning rates? like dou...

Furkan Gözükara SECourses•7/21/25, 5:43 PM

yes

Furkan Gözükara SECourses•7/21/25, 5:43 PM

try square root of batch size

minaricrew_10803•7/22/25, 3:35 AM

Hi! I just downloaded Kohya and all necessary models in my Runpod. How can I use best configs json file(e.g. 48GB_GPU_Quality_Tier_1_40000MB_9.3_Second_IT_Trains_T5_and_T5_Attention) in Kohya?

Should I manually put all the parameters?

minaricrew_10803•7/22/25, 4:38 AM

And I tried this config with L40S, but keep shows "torch.OutOfMemoryError: CUDA out of memory. Tried to allocate 72.00 MiB.
GPU 0 has a total capacity of 44.52 GiB of which 24.25 MiB is free.
Process 165082 has 44.49 GiB memory in use."

250718_inpyo_15_20250722-041950.json6.2KB

Mminaricrew_10803 And I tried this config with L40S, but keep shows "torch.OutOfMemoryError: CUDA ...

Furkan Gözükara SECourses•7/22/25, 11:12 AM

yes use lesser config

Furkan Gözükara SECourses•7/22/25, 11:12 AM

probably your images are not all 1024x1024

minaricrew_10803•7/22/25, 12:27 PM

I just checked and it's all 1024 * 1024. Is there any possibility that I'm facing OOM even though I'm using L40S(And I checked the memory usage right before the training)

Mminaricrew_10803 I just checked and it's all 1024 * 1024. Is there any possibility that I'm facin...

Furkan Gözükara SECourses•7/22/25, 12:31 PM

well it uses 40 GB so you shouldnt

Furkan Gözükara SECourses•7/22/25, 12:31 PM

can you try restart machine and try again?

Furkan Gözükara SECourses•7/22/25, 12:31 PM

but maybe kohya made some updates and now it uses more VRAM

minaricrew_10803•7/22/25, 12:35 PM

I'll try and let you know!

minaricrew_10803•7/22/25, 11:47 PM

@Furkan Gözükara SECourses When I change config to "cache_latents = false / cache_latents_to_disk = false", GPU usage dropped to 36GB. Is there any quality drop if I change it like this?

Mminaricrew_10803 <@205854764540362752> When I change config to "cache_latents = false / cache_lat...

Furkan Gözükara SECourses•7/23/25, 12:53 AM

nope

zadkhamaya•7/23/25, 2:12 AM

Hi @Furkan Gözükara SECourses
Thank you for the dreambooth training tutorial, it really very clear and I successfully manage the training for a person and it performs really well

I have question for other case, if I want to train a person with specific clothing style and also in a specific illustration style, what will be the best approach? Shall I train first the illustration style, then the person with his specific clothing style? The aim is I want to generate this person in consistent clothing style and in consistent illustration style.

Thank you

Zzadkhamaya Hi <@205854764540362752> Thank you for the dreambooth training tutorial, it rea...

Furkan Gözükara SECourses•7/23/25, 2:22 AM

for illustration style use kontext model

Furkan Gözükara SECourses•7/23/25, 2:22 AM

it really does good job

Furkan Gözükara SECourses•7/23/25, 2:23 AM

for clothing if you have pictures of person with that clothing

Furkan Gözükara SECourses•7/23/25, 2:23 AM

that would be great to train

Furkan Gözükara SECourses•7/23/25, 2:23 AM

otherwise you can try to train with cloth pictures

Furkan Gözükara SECourses•7/23/25, 2:23 AM

or traing a clothing model and inpaint images of person

zadkhamaya•7/23/25, 2:52 AM

Ah I see. Previously I already prepare the character with correct clothing and the desired illustration style. But I have only this one image.

I plan to create other pose and other situation using Flux Kontext. Will it perform better if I do the train in one go, it has full body clothing style and also illustration style?

minaricrew_10803•7/23/25, 10:07 AM

Is there best config json file for batch 7? I can only find best config for batch 1 in the folder(Kohya_FLUX_DreamBooth_LoRA_v30)

Zzadkhamaya Ah I see. Previously I already prepare the character with correct clothing and t...

Furkan Gözükara SECourses•7/23/25, 4:18 PM

1 in go is best

Furkan Gözükara SECourses•7/23/25, 4:18 PM

but if classes are matching it can bleed

Mminaricrew_10803 Is there best config json file for batch 7? I can only find best config for batc...

Furkan Gözükara SECourses•7/23/25, 4:18 PM

for lora i recommend multiple gpu

Furkan Gözükara SECourses•7/23/25, 4:18 PM

to speed up

Furkan Gözükara SECourses•7/23/25, 4:18 PM

for lora i dont have batch config but you can use batch size 1

Furkan Gözükara SECourses•7/23/25, 4:18 PM

and set batch size 7

Furkan Gözükara SECourses•7/23/25, 4:18 PM

and multiply learnign rate with

Furkan Gözükara SECourses•7/23/25, 4:19 PM

square root 7 * base learning rate

Furkan Gözükara SECourses•7/23/25, 4:19 PM

however batch size 1 has the best quality

Sommm•7/23/25, 9:09 PM

I am using your Kaggle Kohya ss file to start a kohya ss gradio ui, but it fails by saying no conda environment found, and if i manually start i and try to load a json file forf the configuration, it says something like 'str' has not attribute 'get'? please help

Sommm•7/23/25, 9:15 PM