Will AI-Generated Art Push Us Enter A New Creative Era?
(My own view and it doesn’t represent my employer, family or friends. Thanks)
Updates: Stable Defusion open-sourced its trained model publically on 08/22, 2022. It’s a huge community behind it now. If you search how to use Stable Defusion models on YouTube or Google, you will find enormous resources around it, including how to set up the model to run on your local GPUs.
Since then these underline models have gone through a couple generations and the result only gets more impressive.
Finding Attention-base Transformer
Attention-base Transformer models are basically the backbone model for currently AI-generated Art hype.
Most of you may not know that I am a Machine Learning (or AI) Engineer by profession and a hobby artist by passion (https://www.geoyi.art/ <- my own artist website). In recent months I’ve been working with my colleagues using attention-based transformer models, and they performed excellently on our tasks. It makes me curious enough to start reading and digging into Transformer models a bit more. (I have to admit that I only started and certainly don’t claim I am the expert in this space. But I will not be shy that if you give me the source code of these models, I can do some optimization, debug, train the model, and even fine tune it to perform on specific subjects.)
Not too long ago, I started to run into “text-to-image” and AI-generated art concepts, starting from OpenAI’s Dalle-E (https://openai.com/blog/dall-e/) and Google Brian’s announcement of their Imagegen (https://imagen.research.google/). I’ve seen some fantastic AI-generated artwork from Midjouney (https://www.midjourney.com/home/) and was lucky enough to join its Decord community and have witnessed some impressive art their users create at a second basis.
Stable Diffusion (https://stability.ai/blog/stable-diffusion-announcement), another AI-generated images & art platform’s announcement only ten days ago, has officially moved “text-to-image” to open-sourced space. (Though, I believe their trained model checkpoints are still not publicly available for fine-tuning or transfer-learning yet).
Updates: Stable Defusion open-sourced its trained model publically on 08/22, 2022. It’s a huge community behind it now. If you search how to use Stable Defusion models on YouTube or Google, you will find enormous resources around it, including how to set up the model to run on your local GPUs.
The Four Main Platforms for AI-generated AI
Many platforms are claiming they can do AI Art generation but only mention the following four simply because there are hard-core AI researchers and computer scientists behind it, and also they are doing “text-to-image” that utilize large NLP model and transformer based. If you are looking for something like style transfer or text to video, that’s another conversation or blog.
Back to the four main platforms, Imagen and Dalle-E aren’t “public” available. They provide playground or codebase notebooks that you can use to generate images with your own prompts. Since there is not a huge user base, there is not much to report about them. Only for fun that there is a Twitter account “Weird Dall-E Mini Generations” (@weirddalle, has 1.1 million followers) has some super funny (mostly creepy and disturbing) AI-generated art that’s quite fun to see if you have time. For instance
It’s not the level of AI-generate Art you want to see. Though, to be fair, most of this Dalle-E image on the Twitter account or Reddit were generated by its mini model. It can interpret the text into images still quite impressive. Dalle.E 2 is the second generation model that creates some impressive results. (Link https://towardsdatascience.com/dall-e-2-explained-the-promise-and-limitations-of-a-revolutionary-ai-3faf691be220)
Back to the serious topics, if you use either Stable Diffusion or Midjouney, and see what they can create, it’s really jaw-dropping.
Stable Diffusion is doing some amazing text-to-image that uses language models to create the composition, search related image features and use diffusion to create “seamless” scenes from the prompts. I started to read their research publications and codebase on GitHub (https://github.com/CompVis/stable-diffusion), I will report back if I find something really interesting.
What does AI-generated Art mean to the general public and artists?
Uncertainty creates fear. We don’t know what AI-generated AI means to creative industry and space. Understandably, it creates much fear. Artists already struggle to sell their work and find a new gig, particularly during covid and economic growth slows down. Being an artist who relies on selling artwork for a living can be threatened by these impressive AI models.
But what does that really mean to the Art world and artists?
Many artists I’ve known are pretty optimistic about AI-generated art for different reasons. Many of them referred back to the history when the camera was invented and uncertainties and doubts about film and photographs. It was threatening to traditional artists who held paint brushes and created portraits and realistic landscapes. Did that kill an artist’s livelihood? For sure 100%. But can we deny photography is art itself? I don’t think so. Did that actually change the art world? For sure. But did the camera shape a new art world that we did not anticipate at that time? For sure.
Being an artist is about searching for a niche and community where you and your audiences speak the same languages and share similar emotions. It’s about connection. Developing your own art style takes years or decades. It’s scary that AI-generated art can learn from hundreds of years of art history and create something from quite interesting to impressive. Picasso has a famous quote “Good artists copy(borrow), great artists steal.” That says, as artists, we are all directly or indirectly influenced and inspired by other artists or art forms.
Does that mean AI-generated art can be another source of inspiration? 100% sure. I think AI-generated art from creative prompts will change film, gaming, and visual art dramatically. Why?
Art styles and skills require decades of practicing, muscle memory, skill, and diverse art exposure through multiple years. Now AI-generated art or platforms really lower the barrier to semi or hobby artists entering the art world. Artists can borrow and create new art and art forms on top of AI-generated art. Storytelling, making film, or MTV on top of it. It democratizes art accessibility to the general public. Not too soon you will receive personalized postcards from your friends and family who create them using AI-generated AI. You may make an album cover art for your first music album. Or have artwork attached to every poet you’ve written or get invited to a friend’s cocktail party to see their own AI-generated art? Why not?
The above points touched on what AI-generated art means to the public. But personally, another non-trivial impact I get from being on the MidJourney Discord community that I’ve been exposed to and learned so much art history that I’ve never been before. What is concept art in general? What is cyberpunk? Who is James Gurney (https://jamesgurney.com/)? Who’s Thomas Kirkade (https://www.youtube.com/watch?v=uf4lUiJB1DY)? These that I will never get to come across.
The more I got exposed to AI-generated art. It actually made me appreciate the brilliant minds, creativity, and skills that artists bring. People love artists’ posts work in progress or behind the scenes on social media. Simply seeing something created and burned under your naked eyes is pretty fascinating. We do appreciate some human touches to great work. Does that mean AI-Generated Art Push Us Enter A New Creative Era? I certainly think so. My optimistic self also tells me that the general public’s exposure to massive visual (maybe AI-generated) art will create more gigs and opportunities for artists in the near future.
Adding a little salt to this AI-generated art dish, attached my take and attempt from MidJourney Decord to end the blog. Do you think I can create these myself? No, certainly not. These are 100% compressed from other creative artists through Midjourney.
