Put Yourself Into A Universe Of Art - Using Your Images In Dreambooth Stable Diffusion

[@martinkaiser5263]

Hi and TY for this Vid ! The first vid i see thats in a proper speed to follow ! Cant wait until you reveal the secret artist !!

[@mingeffect]

You're welcome! I just released the video detailing that today.

[@vacc06]

Thanks for sharing! Eager to learn more!

[@NirdeshakRao]

Your are a great tutor, thanks for sharing your knowledge with all. 🙂🙏

[@mingeffect]

So nice of you, thank you.

[@abhinavchhabra4947]

Awsome content, thanks putting these videos. Keep posting!

[@mingeffect]

Thank you :)

[@rikenshah1861]

Thanks a lot!

[@atgaming5318]

amazing, thanks!!!

[@eugeniusvision]

Eric, when you have a chance, could you record a short video about the new fast-DreamBooth collab interface. I think they changed quite a few things compared to this video. Dreambooth area looks differently. Appreciate it.

[@mingeffect]

This does seem to be in order. Sometimes the GitHub updates with daily changes. I’ve got that video on the board for production. Thank you for your suggestion.

[@lloydkeays7035]

Hi, I believe the interface changed quite a bit since your tutorial (which is great) I was able to get it all up and working. But now... How do I simply go back to stable diffusion interface with my trained model directly? I don't seem able to find how to do it without retraining the model. Thanks

[@mingeffect]

Dreambooth produces a session folder on your google drive that will contain the trained model. You can then move that model to any other folder (I create a “model” folder, and then access that model folder when using the standard Automatic1111 colab.

[@eugeniusvision]

@mingeffect with chatGPT gaining so much buzz, do you think we soon will be able to train AI in openAI as well as we do in Stable Diffusion and create our own Avatars there as well? Or can we already do that?

[@mingeffect]

I have a feeling it wouldn’t take much for that to happen. As of yet chat doesn’t generate images (or even random numbers for that matter) but it wouldn’t take much for someone to have it generate code to access the api and produce that elsewhere. As t is, the chat is t connected to the internet for pulling data. I have a feeling they are not wanting to test the waters vs google until they’ve got the appropriate legal issues resolved for such.

[@UnfilteredTrue]

How do you get 169 images of yourself? Do you copy same image multiple times to basically overtrain the model? I a struggling with getting the face right.

[@mingeffect]

Hi, thanks for the question. I have another channel where I am often filmed from different angles in different environments. This is key. Different lighting conditions (just enough where the image is still well-lit but with different ambient conditions), and angles. Think of recording yourself in certain 3d apps. You can start at a low angle looking up at your face and then take photos as you circle around yourself side-to-side. Then you set the camera at a higher level and do the same thing. And then again, until you eventually get photos at a very high angle around your face/head. And different environments--as well as different clothing--will cause subtle lighting and color reflections on your face or the rest of your body. It's easy to get that number of images with this approach. Plus the different styles of clothing-with shots close, medium, and far away will help round out your training images. I hope this gives you a bit more to go on.

[@ak_mits]

I think I have just found what I have been looking for! I tried training on my photos (1 with the same background and the other version with background removed) but still I got basic results. It did not take the face of the photo but rather it seems like it used the whole frame. So the results were all like with the same tshirt of the person because the original photos were with the same tshirt. Do you have any idea to fix it?

[@mingeffect]

The key is to have a variety of angles from closeup to far away. Then comes the prompts themselves. Some of mine (with certain trained models) only produced good results of my face when I also included the word "viking" -go figure. I suppose it has something to do with my beard. Think of other words that might actually tune the render more closely to your face. Maybe add businessman, person, or even throw in a 4-letter random set which will also change the render: something like sdd4