Stable Diffusion LoRA training experiment different network dimension part 3

#stablediffusion #stablediffusiontutorial #stablediffusionai

Stable Diffusion LoRA training experiment different trigger words part 7

Stable Diffusion LoRA training experiment different dataset image resolutions part 6

Stable Diffusion LoRA training experiment different number of dataset images part 5

Stable Diffusion LoRA training experiment different num repeats part 4

Stable Diffusion LoRA training experiment different network dimension part 3

Stable Diffusion LoRA training experiment different batch size part 2

Stable diffusion LoRA training experiment different base model part 1

☕️ Please consider supporting me in Patreon 🍻

👩‍🦰 LoRA model 👇🏻

Welcome, everyone!

In this YouTube video, we will be addressing this problem by providing you with valuable insights on stable diffusion LoRA training. We will specifically be focusing on the use of different network dimensions, and how you can leverage them to achieve stable diffusion and improve your results.

Don't forget to watch until the end of the video to fully understand the impact of different network dimensions on stable diffusion and LoRA training.

Before we dive into today's topic, we want to take a moment to ask you to subscribe to our channel and hit the like button below. Are you ready?

So, let's dive right into Part Three!

[mingtsui]

The final conclusion is different from other suggestion of training LoRA, but after seeing the example, your feasibility is much more higher. Thanks for your effort.

[sarpsomer]

Thanks for the comparison. I would choose version 3 ^__^

[morozig]

Thanks. There is still a lot of unknown space between 64/8 and 128/128. Please also make an experiment for 128/8 128/32 128/64 128/128. And it would be a lot more informative to test target in some unusual environment so that we would be sure that the model didn't overfit.

[DreamBelBee]

I just want to share the settings that worked for me:
Dataset: 20-30
Model: hassanblend 1.4
Vae: stable difussion vae
Unet lr: 1e-4
Text encoder lr: 5e-5
Train repeats: 10
Network dim:128
Network alpha:128
Lr scheduler: constant
Train batch size: 6
Num epoch:15
Clipskip: 2
This generate a good Lora normally at epoch 8, and is versatile with other models

[moonusaco]

Please do a video on learning rate soon. These have been a great help. Thank you.

[TheMilmil]

Can you tell us good conbination of dataset image, like face-shot 20, upper body 10, full body shot 5

[temporaldeicide]

Curious but what if alpha was higher than the dim, or the same? Also what are the recommended settings for a whole person? LIke poses, figure and shapes?

[SH-bjgq]

Thank you for this. My models likeness improved after I saw this and increased network_dim. Though it increases file size.
It's less obvious to me what network_alpha is doing. I've tried it at 1, 32, 64, and 128, and couldn't tell what it was affecting.

Have you played around with the new min_snr_gamma? It really lowers the loss rate fast, and can improve likeness. Possibly speeding up training, as I find it only takes 5-7 epochs to get nice results, where I used to prefer 10.
I've tried it at .0001, .01, .6, 2, 5, 10, and 15 and found 5 was best, as the paper said.

[Leo-bdn]

Hello, I am Korean, but I have a question because I don't understand something.

Conclusion

Choose network dim 128 and network alpha 128 if you do not need to merge with other LoRA models Otherwise choose the same network dim and network alpha as the LoRA models you want to merge with

you said
"We should choose as high as possible if you plan to merge your lower model with other lower models you can use Network Dimension 128 and Alpha 128"


In the PPT content, when merging with other models, the same network dim and network alpha values are required, and when you narrate, when merging with a low-level model, the high network dim and alpha values are entered.

Then, if possible, enter the same network dim and alpha value as the model you want to merge, but when merging with a low-level network dim and alpha model, do I match the network dim and alpha value of the high-level model?

[devnull_]

It is strange you mostly talk about "saturation" when you up the network dimension. Edit - at least in the end you do more observations about this. Also, what was the reasoning behind setting the alpha value? I find it annoying that most of these LoRA and TI tutorials don't give any reasoning, why they chose certain values. I get that it takes a lot of time to train several variations, but at least mention that you are not sure or simply tried some value and it seemed to work if you didn't do extensive testing.