Software Engineering Courses (SECourses)•15mo ago

<@205854764540362752> The extract lora from the finetune only works on the base model I used to crea

@Furkan Gözükara SECourses The extract lora from the finetune only works on the base model I used to create the finetune from. Any image created with a different model using this lora is just static.

How do I fix this?

VigilenceOP•10/18/24, 8:08 PM

VigilenceOP•10/18/24, 8:25 PM

I fixed it, linked to wrong base model.

VigilenceOP•10/18/24, 8:30 PM

@Furkan Gözükara SECourses Extracted lora (left) vs Finetune (right).

VigilenceOP•10/18/24, 8:35 PM

The extracted lora is actually better than the finetune!

VVigilence The extracted lora is actually better than the finetune!

Furkan Gözükara SECourses•10/18/24, 10:02 PM

that means your model overtrained :d

SSiggiSmilez Will you cover how to half the size of the checkpoint too in this tutorial? It w...

Furkan Gözükara SECourses•10/18/24, 10:03 PM

it is ready already : https://www.patreon.com/posts/how-to-convert-114003125

Patreon

How To Convert 24 GB FP16 Fine Tuning / DreamBooth FLUX Checkpoints...

Get more from SECourses: FLUX, Tutorials, Guides, Resources, Training, Scripts on Patreon

VVigilence <@205854764540362752> The extract lora from the finetune only works on the base ...

Furkan Gözükara SECourses•10/18/24, 10:03 PM

different model what you mean?

Furkan Gözükara SECourses•10/18/24, 10:03 PM

which different model

FFurkan Gözükara SECourses which different model

VigilenceOP•10/18/24, 11:08 PM

I picked the wrong base model by accident.

FFurkan Gözükara SECourses that means your model overtrained :d

VigilenceOP•10/18/24, 11:12 PM

I mean the colors are nicer and the textures more pronounced over the trained model when using dev paired with lora.

VVigilence I mean the colors are nicer and the textures more pronounced over the trained mo...

Furkan Gözükara SECourses•10/18/24, 11:14 PM

yes it is very likely model was overtrained and LoRA extraction reduced this impact

VigilenceOP•10/18/24, 11:14 PM

Maybe so, I used the 200th epoch but at 150 the textures are slighly more pronounced.

VVigilence Maybe so, I used the 200th epoch but at 150 the textures are slighly more pronou...

Furkan Gözükara SECourses•10/18/24, 11:15 PM

i also find that epoch good for my 15 images dataset

VigilenceOP•10/18/24, 11:16 PM

Yeah it was nice with image prompt, composition, texture quality, layout and overall amount of detail (items placed nicely).

VigilenceOP•10/18/24, 11:19 PM

@Furkan Gözükara SECourses I asked bmatis and he added a feature to the gui that was in the code, the ability to set the max amount of save-state folders to what ever we want. If set to 1 it will be overwritten with the newest save-state made.

VVigilence <@205854764540362752> I asked bmatis and he added a feature to the gui that was ...

Furkan Gözükara SECourses•10/18/24, 11:22 PM

nice

ApathyJones•10/19/24, 12:44 AM

trying to train on a Flux model but getting this error with it... File "D:\kohya_ss\sd-scripts\library\flux_utils.py", line 81, in analyze_checkpoint_state
max_double_block_index = max(
ValueError: max() arg is an empty sequence

ApathyJones•10/19/24, 12:45 AM

It works fine with the default flux dev, but not this other finetuned dev ccheckpoint

VVigilence <@205854764540362752> I asked bmatis and he added a feature to the gui that was ...

�

𝗡𝗶𝘁𝗶𝗻 𝗣𝗮𝗿𝗲𝘃𝗮•10/19/24, 7:54 AM

Does this mean we can now resume training where we left or if I want to train again with more epochs can I just add it and start from where it last trained?

AApathyJones It works fine with the default flux dev, but not this other finetuned dev ccheck...

Furkan Gözükara SECourses•10/19/24, 10:07 AM

you need to report to kohya

�𝗡𝗶𝘁𝗶𝗻 𝗣𝗮𝗿𝗲𝘃𝗮 Does this mean we can now resume training where we left or if I want to train ag...

Furkan Gözükara SECourses•10/19/24, 10:07 AM

yes you can

Furkan Gözükara SECourses•10/19/24, 10:07 AM

you can also use latest checkpoint

FFurkan Gözükara SECourses nice

sergio_gonzo•10/19/24, 12:43 PM

dr Furkan i hope everything is ok , the full tutorial about make drambooth with flux is in process right?

Ssergio_gonzo dr Furkan i hope everything is ok , the full tutorial about make drambooth with...

Furkan Gözükara SECourses•10/19/24, 12:49 PM

yep editing

Furkan Gözükara SECourses•10/19/24, 12:49 PM

sergio_gonzo•10/19/24, 12:50 PM

how this process make dreambooth models intead of loral model make better models thna make the loras? only to undertand how is work

Ssergio_gonzo how this process make dreambooth models intead of loral model make better models...

Furkan Gözükara SECourses•10/19/24, 12:52 PM

fine tuning is training full model

Furkan Gözükara SECourses•10/19/24, 12:52 PM

lora is optimization

Furkan Gözükara SECourses•10/19/24, 12:52 PM

thus fine tuning better

dxqb•10/19/24, 2:13 PM

@Furkan Gözükara SECourses not sure if kohya and/or you still see this on a closed PR, so here too:

@bmaltais can we add this to gui so many people want to test

Specify a large value for --prior_loss_weight option (not dataset config). We recommend 10-1000.
Set the loss in the training without using the regularization image to be close to the loss in the training using DOP.

doubt that this is a good config recommendation.

all my proof-of-concept samples were at weight 1.

Yes, loss of reg steps is then very low compared to train steps, but this is a function of the reg step prediction already being very close to the target.

you can try at weight 10 or even 1000, but I'd expect that the regularization then overwhelms the training steps, and the model doesn't learn anything anymore.

Ddxqb <@205854764540362752> not sure if kohya and/or you still see this on a closed PR...

Vs12•10/19/24, 2:26 PM

nope it works, just tested it thoroughly, weight 10 works fine just takes longer to train

dxqb•10/19/24, 2:30 PM

even at 1 I think it dampens training of the concept you do want to change, so even if 10 works I'd be careful and/or try different values

dxqb•10/19/24, 2:31 PM

I was more thinking in the other direction, doing less reg steps than 1:1, because even at 1:1 with weight 1, regularisation is strong

dxqb•10/19/24, 2:32 PM

but currently working on sth else so I'd probably not get around to more tests any time soon

Ddxqb <@205854764540362752> not sure if kohya and/or you still see this on a closed PR...

Furkan Gözükara SECourses•10/19/24, 3:10 PM

i am waiting bmaltais to add

dxqb•10/19/24, 3:11 PM

it's merged in sd3 branch

�𝗡𝗶𝘁𝗶𝗻 𝗣𝗮𝗿𝗲𝘃𝗮 Does this mean we can now resume training where we left or if I want to train ag...

VigilenceOP•10/19/24, 3:22 PM

You could always do this if you used the save-state options. You would then paste the path of the folder in the resume from state box. However, the issue is that if you trained for 100 epochs, and the setting was set for 200 epochs, resuming from the trained state of 100 epochs will then train for another 200 epochs, so 300 in total. The epochs in the setting would have to be adjusted if you want max 200.

The new setting is to change the maximum amount of save-state folders you can have. Normally it would save either at the end, which is not great if something bad happens, or at every time you save a copy of the model. This can be a lot of big files, so with this option you can have it set to say a max saved-state amount of 1, so that each new saved-state folder overwrites the older one. Great for saving HD space.

wardensc2•10/21/24, 5:33 AM

Hi @Furkan Gözükara SECourses I try to use your Quality_1_23100MB_14_12_Second_IT.json to train checkpoint but it's very difficult to get 23.1 GB Vram. Windows mostly eat 1.2 - 1.3 GB VRam. How do you limit VRAM use by windows. Since I have 24gb VRAM choosing Quality_1_15500MB_32_Second_IT.json is double the training times and not ultilize my VRAM.

Carlos•10/21/24, 11:58 AM

Hello, I want to try to finetune flux dev on a style rather than a character? Any advice on what is different? For example the captioning, I only captioned with my trigger word + class for the character, is that the same for style? If you have any documentation or training (even if not from you) that's be very much appreciated!! love

CCarlos Hello, I want to try to finetune flux dev on a style rather than a character? An...

VigilenceOP•10/21/24, 4:12 PM

For styles you can either go with a trigger word for the entire image, fully caption everything, or no captions. Characters require captions for everything except the things about the character, like their hair, clothing etc. Things that are captioned can be changed later, things not captioned are static and will be reproduced when prompted.

Wwardensc2 Hi <@205854764540362752> I try to use your Quality_1_23100MB_14_12_Second_IT.jso...

Furkan Gözükara SECourses•10/21/24, 10:09 PM

yes in that case

Furkan Gözükara SECourses•10/21/24, 10:09 PM

just increase number of block swaps

Furkan Gözükara SECourses•10/21/24, 10:09 PM

shown in coming video

CCarlos Hello, I want to try to finetune flux dev on a style rather than a character? An...

Furkan Gözükara SECourses•10/21/24, 10:10 PM

yes i have a full guide and research

Furkan Gözükara SECourses•10/21/24, 10:10 PM

https://huggingface.co/MonsterMMORPG/3D-Cartoon-Style-FLUX