Stable Diffusion v2.0 fine-tuning with DreamBooth on Free Colab

Learn how to fine-tune stable diffusion 2.0 model using Dreambooth. This entire tutorial works fine on Free Tier of Google Colab.

At the moment of testing, this works fine for pictures than styles. Try it out yourself. This tutorial can help you create your own AI Avatars and Ai generated profile pictures for free.

1.5 hrs, Google Colab Tesla T4 for 6000 Steps.

This tutorial teaches end-to-end stable diffusion v2 fine-tuning with Dreambooth starting from Data Collection, Data Processing, Model Training, Model Inference, Converting Diffusers model to ckpt for Automatic1111 compatibility

[DannyCatacora]

Never stop sharing Colabs! People here dont realize they just have to create a conda env and export to a yml that they can just run locally. Great job dude, I appreciate all that you do!

[islamicinterestofficial]

Thanks for the video. But how we can pass our input text prompt against our training image? Means we train the stable diffusion both on our text prompts associated with the specific image?

[CRIMELAB]

Thank you. Im want to create my own model. Very helpful.

[ImanSamanta-eq]

from where did you get class_dir images

[LoneRanger.]

Thank you for sharing your learning experience. It was very helpful.

[neamedia]

base resolution for the SD 2.0 model should be 768x768 - you are using 512x512 - so that could be why your outcome is not perfect...

[TheOfficialCasso]

Hey, could anyone please give me a hand?
20:00, when trying to run the script, I instantly encounter the following error: "/bin/bash: line 1: accelerate: command not found"
I am quite unfamiliar with this so I do not know what should I do to fix that; should I have somehow installed "accelerate" previously? Or what should be done?

Many thanks in advance for your help! :)

[GauravSharma-fkcl]

Will rtx 3060 8gb do the job pls reply broooo❤❤❤❤

[CRIMELAB]

Plz help @1littlecoder can u please explain how do i use my model to generate art after uploading it to Huggingface? I uploaded the the ckpt file, whats the next step so i can start generating? i cant find a tutorial for this.

[geneoverride]

Great video.
Comparison on learning rate, steps, etc would be a great video because not many have done work on this. If you make a video about this, it will catch fire on the internet among SD users. Cheers

[ysy]

Thanks again for creating this video. How difficult will it be to make this work locally rather than on Colab? Anywhere I can get some tutorial ?

[bbibas]

On the last part of the google colab " Run for generating images" is where the i will generate my images, i made several prompts.

if i copy the code and paste with diferent prompts it will generate all my images? so i can generate multiple images with one "play"...

[escappate]

Tesla T4, 15109 MiB, 15109 MiB

ERROR: is not a supported wheel on this platform.
:(

[randomscandinavian]

This is very different from the Fast Dreambooth I have used before but the notebook instructions appears to have just as many unanswered questions. So if I am to train two characters (me and my wife) into one model then I need to add our photos as separate training concepts? And then set train batch size to two or just one? My goal is to have two recognizable people together in one image. Apparently possible but not easy.

[LoneRanger.]

I have been struggling with getting a face trained using dreambooth. I have been told that it’s because of the choice and number of images that I am using. Would you be able to share you learnings on this? How many images, what type, examples perhaps? If you could make a video training faces (for example elon musk, but using a unique prompt to demonstrate it), that would be extremely helpful. Thanks

[puppygozes]

no need of negative prompt for dreambooth on sd2.0?

[LoneRanger.]

Do we HAVE TO convert to ckpt file in order to use it? Can we not use a .pth version (hyperparameter)?

[Mariojav]

Thanks for your help and I have a question, how much GB do you need to use for all these on average on Drive?

[elwii]

How can you use the images from your drive? I copied the path of the drive folder and inserted it for instance_data_dir. But it seems not to be working.

[conanssam]

it's too slow on colab free ver, is it essential colab pro version?