Stable Cascade LORA training

[@Foolsjoker]

If I never hit start without knowing what I was doing, I'd never do anything.

[@FiveBelowFiveUK]

true words. well said.

[@techniquebeats_dennis]

Hey, I was just about to create my first cascade LoRa and was a bit overwhelmed by OneTrainer. Thanks for the video ^^

[@FiveBelowFiveUK]

how did you get on with it?

[@jiaweiwang]

Where did you get that lora loader node? Mine in comfyui looks different than yours. Is that just the baisc lora loader? Or something else that I need to update?

[@FiveBelowFiveUK]

i tend to use the Lora Loader Stack which is part of Crunchyroll custom nodes, you can find it in the manager under custom nodes.

[@FiveBelowFiveUK]

this video featured the Lora Block loader, this is part of the Inspire Pack, also found in the Custom Node Manager.

[@WhySoBroke]

This is very useful! Have you tried training a Lightning model?

[@FiveBelowFiveUK]

i have not yet

[@PlayerGamesOtaku]

Hi, could you tell me how to create Checkpoints for Pony XL on One trainer?

[@FiveBelowFiveUK]

unfortunately i have never tried to fine tune the Pony checkpoint personally, but i do know it requires some special approach, sorry can't say much more beyond that if you train it like the base SDXL, it may not produce a good result.

[@popular75]

you're great.does sdxl lora compatible cascade?

[@FiveBelowFiveUK]

unfortunately it was incompatible layer dimensions, though, i have had success with SDXL loras in cosXL @ 1024x1024 with cascade what i did in my examples was use a normal SDXL txt2img generation flow, with whatever SDXL lora/cnet you want as normal, then use the Cascade img2img with the final image is the input OR cascade txt2image with clipvision input using the SDXL output. This is an "image bridge" so the latents size mismatch won't matter, although you will lose accuracy due to lack of knowledge in cascade.

[@stvn0378]

Pretty great intro! What's the status of using Kohya for this? Would you consider making a kohya tutorial if it's a viable option? =))))

[@FiveBelowFiveUK]

I hope so, they have a large project so it took some time for them to bring cascade support across. I'm circling around a bit so will be covered soon!

[@stvn0378]

@@FiveBelowFiveUK thanks dood, looking forward to it!

[@Sophia-t4b5q]

Suggest adding description version to KZbin video titles, thank you so much for share everything.

[@FiveBelowFiveUK]

yes, it's difficult when we cover things that just released and then they change everyday and technically it's all the same Stable Cascade Version. I thought about adding DAY ONE, DAY TWO etc but that doesn't work either. Strangely people are sent to the early popular videos and then not aware of the newer videos. This created some confusion. I can put links in descriptions, i can add overlay cards and make playlists, but ultimately my hands are ties and people will do what they want :) Thanks for watching

[@hleet]

Thanks for the tutorial ! so this "one trainer" works also for SD1.5 I guess, it's still the same procedure ? 12GB for your cascade lora that's insane ! (I only have 8GB).

[@FiveBelowFiveUK]

It should be working with 1.5 and other models, its very nice for small projects and local training. also they have an active discord community filled with helpful people :)

[@flkadjsfklajfkl]

Followed instructions but when I start training GUI seems stuck at 'loading the model". Cmd shows it's downloading the 6.25 GB pytorch_model_safetensors file

[@flkadjsfklajfkl]

Now it's done it started downloading a 14.4 Gb diffusion_pytorch_model,.safetensors file ! :S

[@FiveBelowFiveUK]

yes those are the models you want to train with being downloaded to your huggingface cache.

[@FiveBelowFiveUK]

cascade has quite a few models required to train afaik

[@flkadjsfklajfkl]

@@FiveBelowFiveUK Okay thanks.

[@SageGoatKing]

Used have bad internet out here in the mountains, now I have Starlink.

[@FiveBelowFiveUK]

i should look into that :)

[@ThePowerfox18]

Thanks! Got my first training in, but it’s like nothing was learnt. Do you have any advice for the parameters for training clothing? Or should I just copy what I did for sdxl, because it worked there (Kohya)?

[@jackyjack2580]

I had the same issue. the default learning rate of 3e-6 is way too low, I tried 3e-4 and it worked great, maybe a bit overtrained though. An ideal learning rate is probably in between those values.

[@ThePowerfox18]

@@jackyjack2580 thanks for the reply! I guessed it was the problem. I did another run and it overtrained on not desired features but didn’t give me the quality kohya on sdxl gave. I’ll definitely try again with your recommendation, but if it won’t work I’ll wait and let the smart people figure it out ;)

[@FiveBelowFiveUK]

the next video will explore a high LR setup, we can take a look together soon :D

[@WhySoBroke]

@@ThePowerfox18if you figure the right parameters please let us know. Have you joined the Discord?

[@byxlettera1452]

With kohya_ss I used to define the keyword that identifies the object you want to train. And when I need it in the image created by the model it was enough to include this particular word in the prompt. I followed your tutorial and I was able to train but I don't know how to specify this keyword in order to include the trained object in the image. Can you help us. Thanks

[@FiveBelowFiveUK]

I made the dataset available for download, this way you can see that each image is paired with a text file of exactly matching name. You describe the image with words inside the text file for each pair. The Keyword you want to use should be inside those lines. iirc you do not need to state the trigger word anywhere in OneTrainer, it can read those captions and your trigger will be inside each of them.

[@byxlettera1452]

Thank you for your help. Your video are the best tutorial I have found so far. High quality material and top topics. I will follow your advices.

[@zhengshen5324]

shall we begin the pixart training?

[@FiveBelowFiveUK]

there might be a preset for onetrainer, iirc

[@SelfErkek-qb8nc]

Thanks

[@luislozano2896]

Why not just install "Stability Matrix" and let it install OneTrainer ? sure you got to teach how to do it the long way!

[@FiveBelowFiveUK]

Nice Promotional Post. Stability Matrix is a wrapper. We cover the original repos from the authors. There are MANY services that take other projects and wrap them into a UI or webpage. This channel is not monetized so time is not a factor, getting the info across is the only concern here ;) i'll say that installing Comfyui/ automatic/ kohyass/ onetrainer "all at once" ? seems silly when you already have them installed on your PC separately.... it's only useful for people that did not already have all of them already. big shrugs.