Catching up with the updates

[FurkanGozukara]

Hello @bmaltais . I hope you are ok. Community is demanding to catching up the updates if possible. Thank you so much for your efforts

Here few list

kohya-ss/sd-scripts@8e378cf
kohya-ss/sd-scripts@bdf9a8c
kohya-ss/sd-scripts@420a180
kohya-ss/sd-scripts@2cb7a6d

These are some of them.

But mostly add the option in gui for huber_scale
Sd3.5 training options for clip l and t5 as well
new schedulers that were added, prodigy plus/ schedule free ones
weighting_scheme options for sd3.5 and flux
block swapping option for sd3.5 as well

[bmaltais]

Yeah, this is something I have put off for a while. Adding SD3 LoRA support will be a pain.

[Sylsatra]

Yes please Progidy with p slice!

[FurkanGozukara]

Yeah, this is something I have put off for a while. Adding SD3 LoRA support will be a pain.

awesome thank you so much

[XCogni]

@bmaltais Huber scale can be used with higher value than 1.0 you can change the max value to 10

[bmaltais]

OK, I have implemented a MVP for SD3 LoRA training in the GUI. Gove it a tray and let me know how it goes.

I don't use SD3 much so I won't be able to properly test the implementation. Please report issues if you find them.

[XCogni]

2024-12-31 07:30:15 INFO clip_l is not included in the checkpoint and clip_l_path is not sd3_utils.py:118
provided

                         getting this error even though i put the path for the clip l model

[bmaltais]

Does it appear in the toml file under clip_l?

[XCogni]

Nope only the t5 and clip g ones show up, clip l does not

[XCogni]

also got a dtype error when trying to train under fp8

[bmaltais]

Let me check why clip_l does not show up... weird... I have been able to reproduce the issue on my system... so I can properly track the issue

[XCogni]

try to get training to start i think

[bmaltais]

OK, I got the training to start. There was a few bugs. Should somehow work now

[bmaltais]

Now I discovered a bug with the config save and load... more fixing needed ;-(

[XCogni]

Also you need to add weighting_scheme uniform as well. Sd3.5 will train better with it

[bmaltais]

OK... added. My issue with config save and load is fixed... Was just a browser GUI refresh issue.

[XCogni]

Training started! thank you for the updates, time to test out the results

[XCogni]

damn sd3.5 is just horrible

[5KilosOfCheese]

damn sd3.5 is just horrible

I actually like many aspects of it. However... Being used to Flux and how it works, it's hard to get used to it. There were certain things I was able to do better with SD3.5 than Flux, and vice versa. But what I liked most about SD3.5 is that it's runs bit faster - especially when changing prompts. At least on Comfy and my computer. However... I still prefer flux. But this is also because I mainly fuck around with my own fine tunes and like to see what I can fine tune.

[XCogni]

@5KilosOfCheese yeah for styles it seems decent to train on, but training for realism, the anatomy and distortions are pretty bad