Google Colaboratory
16 Replies
Ohhhhh, smart. You're using Automatic1111 in Google Collab... makes a ton of sense. If you don't mind, can I ask you the question I asked Dr. Furkan Gözükara because you seem like you have tinkered a bit with this all. Do you think Dreambooth or TI is better for your purposes? I had good success with Dreambooth but the huge file size and having to switch to that ckpt for that specific token felt inefficient. I'm day two into Textual Inversion and getting okay results, but not mind blowing yet. But now I'll be doing it in Automatic so can check what I'm doing against the video more easily.
Yeh this is how I run A1111 all the time since my computer is slow. It has been the most stable but a little costly.
Honestly I am still learning but I think we will need to be skilled at both. I will look to use TI's as much as possible but my dreambooth test today impressed me as I was able to pull my face up even in very long prompts.
All that said i guess im still figuring it out lol
2.1 made a huge improvement on my TI tests, I think it is closer to dreambooth then it was in 1.5 imo
What are you looking to create?
I am looking to create a consistent character for a graphic novel. Kind of a side project. This is where I got kind of stuck with Dreambooth. I suppose I could combine two concepts "character" and "art style" in Dreambooth but that seemed like a much better task for TI. Not to mention all of the various concepts I would want at my disposal to form a visual landscape such as environmental, clothing, hair styles, etc.
Yeh I gotta try a style.
Might be TI for the style and subject since, then you could switch to a dreambooth model you created for cool backgrounds and Img2Img/ inpaint We all will need to develop a workflow once we understand the tools
Might be TI for the style and subject since, then you could switch to a dreambooth model you created for cool backgrounds and Img2Img/ inpaint We all will need to develop a workflow once we understand the tools
Yes definitely. That person's account is incredible. The images are so clean. I love using drawings to prompt pictures, it's like magic. This space is moving so fast I think everything will get easier and better for everyone, at least I hope so. The collaboration between man and machine here has incredible potential.
Yep and he used a custom@model for the backgrounds
Oh so you're suggesting TI for character AND style, then using Dreambooth for BG? Based off of Stelfie's workflow?
Sorry to keep bombarding you with questions. haha. Just nice to talk to a fellow tinkerer.
Just showing an example.
I think individually we will find what works best for our workflows.
I’m just trying it all atm
I think individually we will find what works best for our workflows.
I’m just trying it all atm
Hello @Haashim and @polemistisd'gnosis how would suggest to approach building two people images, for example, "Joe Biden and Donald Trump facing off in the wresting ring", I have tried SD2.1 and other models with face correction, but it does not produce really good results. Any suggestions on what the pipeline can contain?
Hmm, I have never done two people
I would definitely be trying out Dreambooth
Maybe train a Dreambooth model as @Haashim Hasshim said, where you train two separate concepts then merge the model into another model with the same base as @Furkan Gözükara SECourses has shown in his videos like Protogen. That way you could work on two separate subjects with inpainting and call up the concept for each face individually in that model without having to load a different CKPT file each time. But if you're using celebrities, I think TI would be suitable for this as well e.g. "donaldtrumpembedding-1200 facing off with joebidenembedding-1600 in a wrestling ring" . Lots of ways to skin a cat here.
ye sounds like a good plan
Btw I hope you and your family are doing okay right now! Was thinking of you when I saw the news of the earthquake. Sending wishes for safety!
thank you so much. yes Turkey is in very bad situation.
my close relatives have a lot of material damage