Comment by jerojero | Hacker News Reader

Comment by jerojero | Hacker News Reader

There are some fundamental pieces missing in AI art generation that when solved will completely change the game. Forever. I think most fundamentally is the capability for AI to have some kind of memory or maybe more technically a way for the AI to be capable of do style "character style transfer" more effectively. I think this is probably possible, it's akin to making deep fakes but for drawn/photography art.

With this tool suddenly the effort of making a series of compositions that are coherent will dramatically change the game, specially in the videogame industry. I think there are a lot of programmers out there capable of making great games but might be lacking the resources to fully complete their visions due to having to needing assets for their games.

It seems to me like the approach stable diffusion has taken has dramatically increased interest and utility of these tools. So I'm hoping they follow similar lines for other types of AI. Every week I'm reading of a new novel use for these generators that I hadn't really considered before.

It sounds like you're literally describing Dreambooth. You can easily and quickly train models for certain styles like Disney characters, certain illustrators, specific artists and even tailor the model to a specific person so you can produce countless pictures of them in different poses, locations and clothes. This is already available.

Indeed: https://dreambooth.github.io

Note that training is extremely expensive, and is beyond the capabilities of most end users. Here are the details of their training method:

> Given ~3-5 images of a subject we fine tune a text-to-image diffusion in two steps: (a) fine tuning the low-resolution text-to-image model with the input images paired with a text prompt containing a unique identifier and the name of the class the subject belongs to (e.g., "A photo of a [T] dog”), in parallel, we apply a class-specific prior preservation loss, which leverages the semantic prior that the model has on the class and encourages it to generate diverse instances belong to the subject's class by injecting the class name in the text prompt (e.g., "A photo of a dog”). (b) fine-tuning the super resolution components with pairs of low-resolution and high-resolution images taken from our input images set, which enables us to maintain high-fidelity to small details of the subject.

Each fine-tuned model is a copy of the original model. So if the model is 10GB, the fine tuned version will be a separate 10GB file. That might not sound like a lot, but it quickly adds up.

In this case, end users are artists. One could imagine a cloud-based art program which will fine tune on demand. That certainly seems like a good startup idea.

> Note that training is extremely expensive, and is beyond the capabilities of most end users.

You can't run it on consumer hardware, but you can just rent a GPU (or use a free collab book) for a few hours to generate the model. Then you download and reuse it locally at will. Yes, you need storage and if you train often it can get expensive, but it is by no mean out of the end user, at least professional end users, capabilities. And of course there are growing libraries of freely available pretrained models.

> In this case, end users are artists. One could imagine a cloud-based art program which will fine tune on demand. That certainly seems like a good startup idea.

Very much agree about this. At least for a while, I strongly believe that AI will just be another tool for artists willing to embrace it, far from replacing them.

"Extremely expensive" - you can do dreambooth on a free colab with recent optimizations.

This area is moving really fast, one-click solutions are already being created.

People are also averaging weights of multiple models to create new models based on multiple other dreambooth models, and it works surprisingly well.

The stable diffusion reddit is a good place to see how all of this is developing.

Stable Diffusion is probably far from the state of the art, but good $DEITY did it open the floodgates! Just watching the open source scene evolving as a bystander is interesting.

What kind of GPU would you need to run this locally? (the training I mean)

Using offload to CPU (with dreamspeed) one can get away with a 8gb gpu. I haven't tried training myself yet, but there are reports of it working with 8gb vram (here for example https://www.reddit.com/r/StableDiffusion/comments/xwdj79/dre...

I have an rtx 2070 with 8gb and it has been working quite well for me. However there are always models that will not fit. For those running on the cpu with potential nvme offload is not that bad. For example a single inference on bloom 7b (30gb of ram required just for weights) on a 32gb ram machine takes about 30s (it has to offload few GB to nvme). This is on zen 3 ryzen and with no gpu use. I can't wait to try cpus that support avx512.

Dreambooth extension for automatic1111 just came out. Can be run on CPU even. Haven't tried that extension myself yet, but I have followed a youtube video last week and I trained it using a colab in a few minutes (of training - getting the hang of the whole process was maybe 20 or 30 mins including watching the vid?). I think dreambooth is already perfectly within reach of regular users, and is already being used for... ehem... questionable purposes that took 1000's of images and days and days of tweaking and training for 'traditional deepfakes' just 6 months ago. The pace of advancement is breathtaking, it's literally impossible to keep up even if you spend 100% of your time on this.

Thanks for the tip about automatic1111 with dreambooth!

Do you happen to have a link to that YouTube video you followed?

It was actually two now that I look in my bookmarks: https://www.youtube.com/watch?v=w6PTviOCYQY and https://www.youtube.com/watch?v=FaLTztGGueQ . That first one is about 1 month old and was already about running Dreambooth locally. So to be perfectly honest I'm not really sure any more which info I got from which video. Probably most from the second one. It's worth it for just the thumbnail - although I haven't been able to get results at that level yet, holy smokes is it impressive.

Not OP, but this is fastest tutorial for getting dreambooth up and running in automatic1111 I’ve found:

https://youtu.be/_GmGnMO8aGs

I'm building an app that lets users with no technical background train their own dreambooth models for $2-$4: https://synapticpaint.com/dreambooth/info/ They can also share their trained models for others to use. I think making this easy (no figuring out how to do a git pull or rent a gpu) plus the community sharing aspects will make this technology a lot more accessible to artists and general users.

I wonder where it is going to. I was thinking of my favorite concept artists and what it means for them. Probably they can churn out 10x the work, and the role becomes more of curation, as in, scanning and judging the work put out by SD.

They probably still going to adapt the work to fit their needs, and so companies probably still want actual artists for entertainment and artistic purposes.

Small companies (and individuals) probably can use it to circumvent costly stock images all together, so the lower tier photographers/artists there gonna have a problem.

I think most actual work will be done via img2img, outpainting and inpainting, and an real artist can actually give an important contribution.

At the end of the day you still want something that sticks out, and artists are trained at that.

Aren't the fundamental pieces already there?

1 Generate a bunch of characters or objects based on a prompt.

2 Pick the one you like

3 Tell the AI to extract the character's traits from that one picture

4 Miracle happens (I don't know. What does a "character definition file" look like?)

5 Make a new prompt, but add the character definition from step 4, so you get the same traits, only in a different setting or position, etc

This style transfer is possible through both textual inversion and dreambooth models, though they take a while to train.

https://textual-inversion.github.io/

https://dreambooth.github.io/

Training can be done in under an hour[1], really not that long. And yes, what OP is saying is already possible, which seems to be par for the course for this "new" AI space, as it's moving so fast.

[1] https://colab.research.google.com/github/TheLastBen/fast-sta...

It is possible, but right now it still takes quite a bit of time and effort to get it right.

The main challenge is finding the right balance between "make something that looks exactly like this" and "put it in a completely different context". Better similarity equals less flexibility.

For now, the most effective combination will be artist + AI, although it does feel a bit like those that incorporate it in their workflow are helping to dig their own grave.

Do you happen to have any screenshots of what you mean? I’m really curious to see dreambooth’s capabilities in the field, and it sounds like you’ve had experience with some of its pitfalls.

Basically OP says that overfitting is a common pitfall, you don't want to overtrain the model, because then everything will look like your training data, and vice versa with not enough training steps. So it's a bit of a balance. If you search for "dreambooth" on the SD subreddit, you will see a lot of examples of dreambooth results and also some that show overfitted and underfitted results. https://www.reddit.com/r/StableDiffusion/search?q=dreambooth...

"AI" is only copy-pasting elements from a database of existing images. It's effectively just a fancy image search engine.

You can use it to make procedurally-generated art, but it's still obvious what the source material was.

This isn’t true, but it’s 4am, and I regret that I can’t type out a full rebuttal on my phone.

One obvious counterexample is stylegan interpolations. If you interpolate between two images, the midpoint is usually unique — it’s often not obvious what the source material was. (E.g. Gwern’s anime interpolations; sure, they’re anime faces, but from where? “All of danbooru” might as well be “all styles of anime ever created.”)

Maybe someone else can argue the point further.

Go to your favorite image search engine and try to replicate anything AI has generated. It's impossible. The results you get from image search are nowhere near as specific as what the AI produces. It's not even close. The only time image search can compete is when you are highly specific, e.g. "painting of the Mono Lisa", both AI and image search will produce very similar results. But for a generic prompt, image search will come up with nothing that gets even close to the query, while AI can produce a highly specific image.

DreamBooth really should have destroyed all doubt about this point point, as the AI can generated highly specific images of subjects that aren't even in the original training set.

> as the AI can generated highly specific images of subjects that aren't even in the original training set

I haven't seen any concrete examples of this yet.

It's pretty impressive that Stable Diffusion can compress over 200TB of already previously compressed images from LAION 5B down to a few gigabytes! And that it can search that database so quickly to copy and paste things from the right images! /s

There's a lot of good discussion to be had around the ethics of AI art, training on copyrighted materials, etc. But it is equivocally not just copy pasting from a database of images.

Yes, neural network compression is, indeed, very impressive.

> But it is equivocally not just copy pasting from a database of images.

Just because you put a lossy compression step in the middle doesn't mean it's not copy-paste anymore.

Reply on news.ycombinator.com