AI Generated Images Are Getting Too Real

LoRA, Model Mixing, and more text to image magic. Now you can generate 5 fingers easily, extreme photorealistic people and illustrations that you would think it's taken from Pixiv.

LoRA: Low-Rank Adaptation of Large Language Models

Koiboi's Video

Special thanks to ButterCream, GenkiKun, koiboi, Niaka

This video is supported by the kind Patrons & YouTube Members:
🙏Andrew Lescelius, Chris LeDoux, Alex Maurice, Panther Modern, Jake Disco, Demilson Quintao, Tony Jimenez, Shuhong Chen, Hongbo Men, happi nyuu nyaa, Carol Lo, Mose Sakashita, Miguel, Bandera, Gennaro Schiano, gunwoo, Ravid Freedman, Mert Seftali, Mrityunjay, Richárd Nagyfi, Timo Steiner, Henrik G Sundt, projectAnthony

[Music] Slipstream

[bycloudAI]

Please don't be weird in the comments 😭
and a new automated merging method just came out right when I finished this video... These technologies are moving too fast man

[themassagechair]

This shit is improving so fast that you can legitimately say "ever since back in January" like it is ancient history.
January was a month ago. This happened in a month.

[DJVARAO]

I do photography, and I have been working with SD since the beginning. I can tell you that the photorealism you can achieve now is unprecedented. I can't even believe that some of the images were artificially made, with face portraits including pore details.

[pathaleyguitar]

This is completely insane. What a time to be alive.

[Sphnxfr]

As an artist who's been following these developments while trying to be honest about the implications for whether I'll be able to turn my little hobby into a career, I have a few comments.
1) the ability to imitate any artstyle flawlessly based off just a few examples is impressive and frightening in its own right. how's anyone going to make a living developing a visual style when that kind of originality takes years of work and anyone can just take those results, cross-breed them with something else and use the design work you put in without ever paying you a penny?
2) the idea that you're going to be replaced 'not by AI but by a person using an AI' sort of obscures the point that it will be far fewer people doing the 'same' work in a highly automated fashion. I think it's safe to say most working artists won't be able to re-train and keep their jobs.

[Phobos]

A few corrections. We have been doing AI model merges/interpolations since at least early 2019, we just used exactly the same for Stable Diffusion. Infinite merging is not sustainable, as the models end up being just a plain average mess. Block weighted merge fixed more than a few problems of regular merging in that you can now control what parts of one model is to be transferred to the other, without affecting other things in the model, that part is certainly recent. You don't really want to "merge" LoRAs into models, you can just use them on the fly with any models, even multiple of them simultaneously. Merging the LoRAs is going back to Dreambooth, where you lose all control.

[Lexsed]

04:26 - my man offered a timestamp to skip a 60 second explanation, time is not literally money for youtube viewers.

All of your videos are so clear and concise that I cannot even imagine anyone wishing to skip any second of this.

[Etrema_YBT]

I remember couple months ago I was sure AI-generated images are just waste of time and will be only temporarily trend once people get tired of making arts with horrible fingers etc. until couple days ago while I was browsing pixiv I stumbled across some really beautifully "drawn" anime characters and backgrounds. Right before pressing "follow" the artist button, I noticed that it was all AI-generated. My jaw dropped. There were no broken fingers, shading, proportions, lights everything was perfect. I'm not an artist, so I thought that I couldn't differentiate because of my lack of knowledge and started googling/youtube searching about people's reactions on recent AI-arts. After that it was clear that even people working in the field of art for years are impressed as well. Dozens of dozens comments about how AI-arts are becoming more realistic and better.

It's really impressive and scary at the same time. I wonder what will happen in the couple of months or maybe 2-3 years time. People who didn't even held a pencil in their lives could make high quality images on their own. Sure, those images may not have meaning/"soul" the authors put in their creations... but the fact that, the person making AI-art will get darn good image is for sure

[YVZSTUDIOS]

Quick note about LoRa: When it was new, it didn't produce good results so it was almost abondoned by the community, but after some time, an updated version came along and from the stuff I've seen, it got exponentially better. The biggest gate-opener tho was that it was usable for ppl with a regular 8GB VRAM GPU (while dreambooth required a high VRAM GPU or Google Colab). I think I gotta give LoRa another try to see it for myself.

[maiLman__]

It's great to be watching pandora's box be opened in real time! Can't wait to witness the chaos this will bring, both good and bad!

[sylas]

I'm a developer and I was inspired into learning and practicing classical art, digital and miniature painting; and tomorrow maybe something else like music. AI generate art is pretty amazing and depressing at the same time. Lots of people might lose their jobs just because of short-sighted, impatience people that will only goes way beyond for fast money and automation. And art will perhaps lost a bit of its meaning is to enjoy the beauty and the hard works to produce it.

[RSK]

AI: "do you know what you want?"
Humans "No! Please tell me!"

[taihuynhuc]

Your video is good and pretty accurate, but there are some small errors tho:
3:50 Sd models can be pruned and converted to fp16. This way the size is reduced from 7.7GB to 2GB, the differences are also hard to notice.

4:13 The size of LoRA you showed isn’t wrong since ppl mostly make 128 dims (ranks). But imo 128 dims is overkill for many cases. You can still get good results with just 64 or 32 dims (about 80MB and 42MB in size iirc). Some ppl even have gotten decent results when training 8 dim LoRA

5:38 Mixing models can’t solve these issues because the source of the problem is the fact that SD is a Latent Diffusion Model. It process latents which are 8x compressed version of RGB images so many small details have been lost. The only way to fix them completely is using t2i models that process RGB image directly like Google’s Imagen. StablitityAI will release DeepFloyd which is a open-source Imagen model soon, so I think they will be soon fixed tho.

[gheshue]

Just wait a few months until we can also animate them, voice generation AI good scary good too, and ChatGPT like language models can give them responsiveness. By the end of the year we will have fully interactive virtual AI that looks real.

[bakapotatoes]

as an artist, this is cool .. also terrifying

[xdragonk]

The scariest of AI generated stuff is that in 1 month, it can out generate the artworks of someone's lifetime.

[GELTONZ]

So I stumbled across one of the AI generated "cosplayers", found out what they used to make it...and immediately installed Stable Diffusion and CoolMix. Over the course of a night I managed to figure out how to turn artwork of any videogame or anime character into a photo of a real person in cosplay. I've spent the past week doing this. It's wild. I might need to make a tutorial video on it because it's amazing seeing obscure game characters no one would ever cosplay come to life. Like Nico Southey from Azure Dreams and stuff like that.

[IkeMonk]

Me as an artist saying this won't replace me, last week... Yeah, this situation just keeps getting worse and worse. And it kills me, that it had built off this sweat and hard work of basically every artist who ever posted on the internet ever. I don't care what anyone else says, this was theft on a monumental scale, and there probably isn't away to put it back in the box.

[bakurascoffeeshop]

AI voices, AI Images, AI chatbots... at this rate you could fake an entire social media presence... or replace real ones

[bigpapa]

After 10 years, if you still have your art and drawing skills done by hands, you will be like a treasure.