Some feedback on workflow:

  - Automatic1111 outpainting works well but you need to enable the outpainting script. I would recommend Outpainting MK2. What the author did was just resize with fill which doesn't do any diffusion on the outpainted sections.
  - There are much better resizing workflows, at a minumum I would recommend using the "SD Upscale Script". However you can get great results by resizing the image to high-res (4-8k) using lanczos then using inpainting to manually diffuse the image at a much higher resolution with prompt control. In this case "SD Upscale" is fine but the inpaint based upscale works well with complex compositions.
  - When training I would typically recommend to keep the background. This allows for a more versitile finetuned model.
  - You can get a lot more control of final output by using ControlNet. This is especially great if you have illustration skills. But it is also great to generate varitions in a different style but keep the composition and details. In this case you could have taken a portrait photo of the subject and used ControlNet to adjust the style (without and finetuning required).

I love how much work went into this.

There's a great deal of pushback against AI art from the wider online art community at the moment, a lot of which is motivated by a sense of unfairness: if you're not going to put in the time and effort, why do you deserve to create such high equality imagery?

(I do not share this opinion myself, but it's something I've seen a lot)

This is another great counter-example showing how much work it takes to get the best, deliberate results out of these tools.

I've done so much with a fine-tuned model of my dog.

I previously made coloring pages for my daughter of our dog as an astronaut, wild west sheriff, etc. They're the first pages she ever "colored," which was pretty special for us. Currently I'm working on making her into every type of Pokemon, just for fun.

I liked the original more than the final version. The vector style drawing was much more futuristic and more interesting.

Seems like lots of work went into that and I hope the author enjoyed the process and enjoys the final result.

I did something loosely related. As a present for my girlfriend's birthday, I made her a "90s website" with AI portraits of her dog: https://simoninman.github.io/

It wasn't actually particularly hard - I used a Colab notebook on the free tier to fine-tune the model, and even got chatGPT to write some of the prompts.

He mentions the Colab for Dreambooth, that only takes ten minutes or so to train using an A100 (the premium GPU) and you can have it turn off after it finishes, and saves to Google Drive. Super easy.

I did the exact same thing when I saw DreamBooth for the first time! I showed it to a bunch of friends and they convinced me to turn it into an iOS app. https://apps.apple.com/app/ai-avatar-for-dogs-floof-ai/id165...

People have been sending me the cute pics the AI generates of their pups. I think this is arguably the best thing so far in this latest wave of AI releases!

The dog picture is really nice, but then it’s hung on the same wall as 20 other crowded pieces of (in my opinion) dubious quality.

This would have been much better standalone.

Valuable processing info in the comments. But why so much effort to produce something without the option to have ownership (copyright) over my product? If I draw a strange line with any digital painting tool and put a circle and a square around, I sign this art and this is my Art. If a spend a day with prompting, upscaling, fixing with Control net in the end of the day I will have a funny picture which is not mine.

https://fortune.com/2023/02/23/no-copyright-images-made-ai-a...

It's unfortunate a lot of the nice artsy detail disappeared when he had to recreate part of the head, but I guess that is inevitable. Great work and interesting writeup.

If anyone wants to try Dreambooth online, I made a free website for this: https://trainengine.ai

I would highly recommend using Photoroom's background removal tool. Does a far, far better job than Photoshop.

But why pick a dog as an example?

Humans are much worse in telling dogs apart than other humans (except perhaps the owner of the particular dog).

So for all we know, the AI didn't generate a portrait of this particular dog but instead a generic picture of this breed of dog.

There might be a few things Draw Things missing from this article: no mask blur, not selecting the inpainting model for inpainting work.

Tomorrow's release should contain both mask blur and inpainting ControlNet, which might help these use cases.

>I was wrong . it seemed to take the top and bottom-most row of pixels and extend them down from 512px tall to 1344px tall.

I mean you cannot outpaint in the img2img tab, load the image in the inpaint tab and possibly use the inpainting model.

Pretty cool stuff. Personally though, not a huge fan of his “the one” choice. Some of the other images in his assortment were much better imo. Each to their own of course though!

Nicely done. I built a t-shirt/mug/frame printing app. I am using stablediffusion (intructpix2pix for selfies) with prompts pulled in from Lexica. The larger images are created with swinir and physical printing is from the good folks at printful.Com.

Big props to folks at replicate.Com for making solid infrastructure for ml.

https://www.ai-ink.me/

Awesome work. I build an app to train dreambooth model and generate images of hich makes this process very easy.

The app also has a rest endpoint for anybody to create app using it. Lot of clients create niche websites catering to different use cases. There is a kind of gold rush going on in this area.

https://aipaintr.com

Also how much $s were spent on the project?

Fantastic! definitely bookmarking. I spent a big part of the last few days attempting this, my model didn't come out nearly so well. I decided that it was because I don't have enough training images, and so have been taking 3x as many pictures of my dog to compensate.

Recently saw a nice replica of Duchamp's work 'Bottle Rack' from 1959. Readymade, but maybe a bit expensive. For the price they asked, I could do it myself in a blacksmithing class and have more fun.

Isn’t disappointing that nothing important is open source these days in AI?

I used Stable Diffusion Dreambooth to generate my github profil picture, what do you think ? https://github.com/dezmou

This is the barely a full step from; "I used Stable Diffusion and Dreambooth to create nudes of a person I know".

Yes, They're Real and They're Spectacular.

This is a great writeup on some of the nuances and gotchas you have to watch out for when finetuning using dreambooth and the generative creative process in general.

It's impressive, The end result was beautiful. I always use to wonder how to generate some meaningful art

Any references where the same has been tried on humans ?

I find it interesting that this green/orange colour palette so commonly appears in midjourney images, seemingly regardless of the subject.

This is really interesting. I do wish the author included the cost to train the model from replicate though.

If I wanted to do this, what kind of specs would I need to have on my Desktop Computer?

Is the link an AI generated tutorial? Write me a blog post tutorial pretend you are …

"art portrait" seems grammatically wrong...?

Results at the top of your article/project please

Great work writing up the process. Much appreciated!

The style looks Andy Warhol to me

What are the tools we can run on a Linux machine?

EDIT: four downvotes and zero answers how to run it on a Linux machine…

[flagged]

Personally I paid a friend $200 to create an art portrait of my dog.

the lengths techbros will go for in order to avoid paying an artist for artwork

as well as, doing all that nn/ml stuff, instead of just, trying to learn a bit of how to make an artwork themselves, how to draw something, even by tracing over a photo, like doing a 'how to do a vector colorful painting dog' search and going off on that.

like, this end result doesn't even look far off from what a 'colorful vector dog portrait' tutorial would yield. it just involves tons and tons of questionably sourced artwork, and violated copyrights. (i know techbros are very confused about copyrights, but stuff like licenses and copyrights actually do have their meanings, limitations, and liabilities)

specifically picking stablediffusion, probably the most blatantly stolen artwork-based model (given how open and clear it is with what data has been used for it, and how you can't just squirm 'i didn't know what were the terms of use of their data' with other, more closed-off services), that's just another great touch as well.