Please re-add the option to use classes / instances (old method)

[hypervoxel]

Please re-add the option to use the previous method of classes / instances (old method). I got better results that this session / instances method

[ghost]

I agree with this. I've run some experiments myself comparing the old and new method, followed the instructions perfectly. Used the same dataset, same number of steps, same model. Labeled my files correctly, with the token and a number, (token (1), token (2) and so on. Token was an entirely unknown, password like string. I'm specifying this because creator likes to blame bad results on following the instructions incorrectly.

New method has a more vague understanding of the character, old method can actually copy some ones likeness very closely. A lot of TheLastBen's examples were ran on characters or actors that Stable Diffusion already had prior knowledge of, aiding in the training process. To truly test it, he should train on an entirely unknown character, a pop star of the 2020s perhaps. And train with the old method and the new one, and make the both the models public. Then, maybe the creator and everyone else will finally see, the new method is worse.

Please, if anyone can back us up here! The more people agreeing, more likelihood we can get the old method back.

I'd also like if he could implement the option to use the old 1.4 model.

To TheLastBen, please don't take this personally or as an attack as you tended to do in the reddit thread. It's just constructive criticism. As a developer, you should be able to handle that without shooting down everyone who tries to help the project.

[TheLastBen]

A lot of TheLastBen's examples were ran on characters or actors that Stable Diffusion already had prior knowledge of, aiding in the training process.

That's completely wrong, training with a different instance name makes it a completely new character for stable diffusion. I bet even if you train on a known actor, you won't succeed

[ghost]

A lot of TheLastBen's examples were ran on characters or actors that Stable Diffusion already had prior knowledge of, aiding in the training process.

That's completely wrong, training with a different instance name makes it a completely new character for stable diffusion. I bet even if you train on a known actor, you won't succeed

Can you identify the reason as to why I am not succeeding? Did you identify an issue within my workflow?

If it's wrong, you could easily prove it. Just train a model on a completely unknown concept, yourself or something like that. Some pop star or actor who became known well after the dataset used for SD was created.

I trained models on myself, and several actors/actresses. I never once got good results with the new method. The character was always vauge and distorted.

Just started a new old method model, trained for HALF the amount of steps. 80% text encoder, class woman, 200 class images. Jaw dropping results. Looks like the person and the prompt dosen't require ridiculous amounts of engineering to get a good image.

[TheLastBen]

@hypervoxel
https://colab.research.google.com/github/TheLastBen/fast-stable-diffusion/blob/main/Dreambooth/fast_DreamBooth-Old-Method.ipynb

[TheLastBen]

start with 2000 steps UNet and 350 steps text_enc, and make sure you use only one word for the images names a word that doesn't exist like "otiittre"

[TetraOrb]

Yes, I use one unexisting word. I'll try again, thank you! =)

[TetraOrb]

@TheLastBen tested the training today. Faces become stylized after 5000 steps. Really love the outcome! Thank you for al your work!

[dvelm]

What does it mean stylized? Does it mean that the face appears everytime in any prompt? Or it means that when the identifier word is written the outputs show the faces correctly?

[TetraOrb]

What does it mean stylized? Does it mean that the face appears everytime in any prompt? Or it means that when the identifier word is written the outputs show the faces correctly?

I try ro teach SD to draw in my style. And on lower Unet steps it tends to make painted people with weird realistic faces like this:

But the more Unet I train, the better the result is for the faces. But I haven't tried training with reference images yet.