When Androids dream: An experiment in using AI to imagine “antiques.”

With the exception of screenshots/training images, every image you see here was generated by artificial intelligence.

---

Nothing is real.
Reality is a dream.
Everything is imagined.

---

This stunning image of a tram crisscrossing its way across a neon lit cityscape evokes imagery of the dark, dystopian world of sci-fi thrillers. Neon lights reflecting on the wet pavement as fares huddle off to the side, awaiting their ride. You can almost hear the dull hum of the city, smell the rain, and perhaps even catch a shiver of the cooler temperatures. And yet, none of this is real. This city does not exist. The trams are merely an echo taken from the dreams of an algorithm.

The majority of us probably think of AI in the context of movies; i.e. the type in which the planet is destroyed by some nefarious software entity that has suddenly become sentient. And who knows? Perhaps our future holds such terrors. Or perhaps not. Only time will tell. But in present day, artificial intelligence is a bit more mundane. From AI powered software designed to discover new medicines, to polite but annoying customer service chat bots, to complex algorithms used to model weather patterns, AI is already here. And here to stay. So it is no surprise that that AI has also arrived in the field of art, graphics, design, and illustration.

What are AI image generators?

AI image generators are machine learning, text-to-image modelers which generate digital images from a natural language descriptions. In simple terms, this means it creates an image from a text description you provide.

There are quite a few out there at the moment, with names like “Stable Diffusion” or “Midjourney,” and needless to say, each comes with its own own strengths and weaknesses in terms of its capabilities and resource requirements in order to create images. Some work with realistic images better. Others excel at graphics and producing logos.

Nevertheless the results are often striking and impressive, to say the very least.

• 

• 

• 

Setting aside installation and hardware requirements (which can be substantial), once you have the backend scripts installed and running, most have graphical interfaces which allow you to enter in your requirements for the image to be created rather than typing in a series of complex code commands.

These controls allow you to define things such as the overall image quality, how much processing power to devote to creating the image (more processing power may yield better results but may take longer to produce), how closely the AI attempts to follow your instructions (vs generalization with more freedom to interpret) and a few other complex variables.

And of course there is the option to define the overall size of the image in pixels in which, again, comes at cost of processing power.

Tweaking each command will effect the final image, so even a minor adjustment or change to just one of the instructions can effect vastly different end results. In general though, the primary language based command in which you describe the image is the most import one. This means descriptive language but language that the algorithm can understand. We will talk more about this later on.

---

---

Depending on the commands entered, one can also blend styles of various artists or even apply these artists styles to particular pieces. For example, one might describe the image they wish to create and then add at the end “in the style of ‘such and such’ artist”.

The examples above are blends of the styles of three different well know Japanese artists. Remember – each of these images is a new creation. These are not copies or blends of existing works but rather new images inspired by a particular image style, artist and/or a blend of both. Nor is this limited to paintings or traditional artworks. You can apply this concept to specific items such as the chair below.

---

---

The amount of time it takes to generate an image can take a mere few minutes or as long as hours, depending on the system you are running the AI on. In this case, because I lacked the processing power and proper graphics card needed to generate images in a reasonable amount of time, I opted to rent space on a server and process the images there in a more expedient manner. In other words, each one of these images costs a small amount of money to generate. However, having found an AI running on inexpensive hosting, the cost was negligible. This makes a huge difference in image quality as not all accessible AI image generators allow access to the full range of commands. Running your prompts on “dumbed down” version will produce less than stellar results.

---

---

So the overarching premise is simple enough:

In plain language, enter in a description of what you want to see and the AI will offer its own “interpretation.”

It sounds simple enough. And at times it can be.

Want to see a painting of a cute dog wearing a hat sitting on an antique style chair? Sure. We got you covered. How about some Tang dynasty style fat ladies having martinis on a boat. We got that as well!

• 

• 

But what about images that are truly “antique” in nature? Can we create those sort of images as well?

I have always wanted to see what a “woodblock print of a fat goat sleeping on an antique Chinese opium bed” would look like. Maybe “two fat shiba-inu dogs sitting on Qing dynasty style bed?

• 

• 

Well, now we know what that looks like! 🤣

But what about furniture… in fact, what about actual antiques? Could an AI imagine “Chinese antiques?”

The idea that one can create an original, never before seen “antique” that is neither a copy nor a reproduction is fascinating.

Intriguing.

First Attempts

My first few results were intriguing but not very attractive. I used phrases such as “A Chinese antique chair” or “A realistic image of a Chinese Ming dynasty cabinet.” And while the items are relatively balanced they are not particularly believable or aesthetically pleasing. Images were plastic, and in most cases unoriginal and even cliche’ish. Some simply unattractive failures.

Not particularly convincing.

And these were the “better” examples. The worst ones I discarded immediately.

• 

• 

• 

• 

• 

• 

The reason for such lackluster results is partially because the AI doesn’t seem to know the names of any classical forms beyond the most simple forms like “bed, chair or table.” Nor does it know any specific Chinese antique terms like “Shanxi” or “8 immortals.” So while it can determine “bed” or “chair”, it doesn’t seem to be able to differentiate between a “canopy bed” vs an “opium bed.”

A look at the image training set available and one quickly understand why. In some areas, there is a lot of quality information and details which it understands. Porcelain is a good example. In others, the data is muddied and make less “sense” to the AI. Furniture seems to be one of those less well understood areas. Most of the descriptions of furniture items overly simplified, lack any real detail and some times are totally inaccurate.

The solution was something I stumbled upon accidentally.

In the example below, I entered as the prompt “a fat man wearing a funny hat standing next to a table with two chairs inside of a Tibetan style home.” However this time, I provided an image of a Tibetan style home from a random google search as a sample image to work from. The results were somewhat unpredictable and what I got in return was these potato people. However, the overall details were correct. Bright colors, a table with chairs, hats and most importantly, a somewhat Tibetan style room.

Here is another, running a slightly modified version of the same prompt. More potato people, but still fascinating. Faces are a problem for the AI in some circumstances, although there are scripts which offer corrections. With the corrections applied, the images are very convincing (however I am not running that particular correction script, therefore the result is potatoes).

Do AI’s have a sense of humor?

Ever seen imperial duck-dogs? The image on the left is of Qianlong Garden in the Forbidden City. This was the training image provided to the AI. The text prompt given was: a cute dog wearing a hat, standing next to 2 ducks and a cow inside a Chinese palace.” In this case we can see the AI combined dog, duck and cow. The duck 2nd from the lower right seems to be a hybrid of a dog-cow-duck. And dog with the hat has cow ears. This was not the last time the AI produced some humorous results.

Not every attempt was a success and, as we see below, at times the results are downright hilarious.

Porcelain

Having “less than stellar luck” with furniture, I decided instead to try my hand with porcelain (at times with a training image as a starting point) using slightly more descriptive terms. The resulting images were interesting, to say the least. Many looked fairly realistic and some were even convincing. But why?

It appears that the data set available to the AI for porcelain is vastly more established, with much clearer imagery of what the various forms are meant to look like as well as basic terms like “Ming blue and white.” Therefore in a relatively short amount of time, I was able to produce somewhat realistic looking images right away on the first few tries.

This was encouraging.

• 

• 

• 

• 

• 

• 

• 

• 

• 

Unfortunately, I did not record the prompts for most of the porcelain pieces. Some likely terms included “Kangxi, Ming blue and white” and even “Japanese Imari.”

And once again, some of the results were absolutely hilarious.

One prompt was “Donald Trump painted on blue and white Ming style Chinese porcelain.” The other was “a Chinese Qing dynasty style porcelain plate depicting a bicycle in the center.” The results were amusing but expected. However, on the former, notice how the AI blends the concept of “Donald Trump” with “Chinese.” Literally. So the end result is a “Chinese Donald Trump.” This happened a few times with faces, and in one instance even a dog ended up getting “Asian eyes.”

• 

• 

• 

Porcelain Patterns

I also thus realized I could create porcelain patterns of pretty much anything I could imagine. What a rabbit hole!

Endless fun. 😃😃😃

• 

• 

• 

• 

Furniture: second attempts

Emboldened by a partial success in depicting “somewhat realistic looking” porcelain and now armed with a better knowledge of how to construct commands, I set out once more. This time, in addition to describing the piece that I wanted, I also described a level of detail, colors, paintings, decoration styles and even the desired lighting. The prompts shown here are truncated as otherwise the descriptions would be way too long.

But here is an example of one prompt in its entirety:

Several pieces of Chinese antique furniture accompanied by Kanxi blue and white porcelain vases, carefully arranged, antiquities, Ming dynasty, classy, elegant, auction house display, Full-HD, hard-Wood, Sotheby's, professional photographer, well lit, octane render.

The results were awesome.

Of course, while not every detail was perfect, the images it was now creating were somewhat realistic, true to form and, although whimsical, balanced enough to catch ones attention. Brilliant. Interesting.

But still not perfect.

Two problems remained.

1. Proportions & balance;
2. Getting pieces to remain “in frame.”

Lets look at both issues…

---

---

Proportions & Balance

The difficulty with Chinese antique furniture is often proportions, with the AI at a loss as to how to balance such subtle delicate proportions, with the end result being wonky overhangs, outcrops and odd connections.

So it seemed either pieces was out of frame; or if they were in frame they were wonky, unbalanced and lacking the necessary symmetry to convey the essence of Chinese Antique furniture.

Interesting, but not particularly pleasing results. If anything, they began to look like something from a drug or alcohol induced “antiques” hallucination where nothing fits as it should be.

Getting pieces to remain in frame

Other times, the images turned out great, with the AI generating rich, detailed closeups of intriguing pseudo antique pieces. But only partially in frame and incomplete. These images leave you wanting for more.

• 

• 

• 

The solution

The solution to this turned out to (again), be in the form of a training image. While adding additional language to the prompt describing proportions did help to some degree, providing the training image was the most effective, as we can see in the images below. Despite these being two different styles, from two different countries, two different forms and two different descriptions, the AI nevertheless is able to grasp the concept of “balance and proportion,” at least to some degree.

Here is another example.

In this example, it does a pretty good job with proportions but it has trouble with the white space from the opening in the front of the example image. Its attempt to interpret this is awkward. Nevertheless the effort is not bad, with the result being a mostly balanced, true to form piece.

Finally, I was able to create something that looked almost convincing, realistic, and in some cases even beautiful. Something that, as a professional, I would add to my collection. Something I would buy. Something I would sell. Something I would show to a customer.

This was also when I recognized the addictive element in it. From crafting your prompt to the anticipation of the outcome, each instruction offers up a hit of dopamine, as you eagerly await the results of what will eventually display. What new undiscovered design will it produce? What colors? Will it be ascetically pleasing? And in true gamblers form I rolled the dice quite a few times. The results were exhilarating. Beautiful imagined designs with crazy patterns, unusual pairings of various styles and in some cases interesting quirks (such as the cabinet with 2 door cabinet with an extra side door)

---

---

Would you buy any of these items if they came up for auction? What about the trunk above? I would. Or perhaps I would bid on the incense stand below.

• 

• 

If they came up for sale, which piece would you consider adding to your collection?

Below are a few of the more interesting examples from Canton style couches to Ming Huanghuali furniture. Each one is unique. Some follow styles more closely, while others are more lose “artistic” interpretations.

All of them are fascinating.

• 

• 

• 

• 

• 

• 

• 

• 

• 

• 

• 

• 

• 

• 

• 

• 

• 

• 

• 

• 

• 

• 

• 

• 

What about Tibetan?

Now that I had managed to created somewhat convincing facsimiles of antiques, it seemed odd to stop at Chinese furniture. What about less refined and more “exotic” or “rustic” styles? Would the algorithm be able to interpret these styles as well?

The answer was a solid yes!

And spectacularly so too!

This antique Tibetan leather trunk turned out quite well. Photo realistic. Convincing. Exotic looking, leaving the viewer to wonder just where it comes from. A blend of familiar styles that could easily come from a region you have purchased from in the past.

Let’s look at a few more examples:

Tibetan items in general, particularly those in which I started with a training image, produced fairly excellent results. This is assuming that terms like “exotic” and “hand painted” are ones the AI can easily interpret. I would love to have some of these pieces in my collection.

If they were real, that is.

• 

• 

• 

• 

• 

• 

• 

• 

• 

Again, at this point, why stop?

What else could I attempt to imagine?

Chinese arts and scholarly objects

Each of these areas is a tricky new adventure, requiring new prompts, new descriptive language and new training images. Essentially it’s starting over from square one. But hey, why not? Look at the results!

• 

• 

• 

Some of these are convincing. I would buy the carved marble panel below. I would buy the brush pot. I would even buy the table screen.

If it existed, that is. 😎

• 

• 

• 

Carved lacquer was tricky, not only because I didn’t have the language to describe specifically what I wanted, but also described in a manner in which the AI would understand. In this case, many of the terms were too specialized with not enough specific information in the data set. Or too much processing time required to create the results. In the end, the better results were when I allowed the AI some freedom.

Fun, but not totally convincing. Or perhaps I should say not convincing to those in the antiquities world. For the uninitiated, some of these pieces look perfectly normal and indistinguishable from the real thing.

• 

• 

• 

Metalwork presents many of the same challenges. These (in some cases) are very complex objects which require a lot of carefully crafted descriptions in language the AI understands. Having crafted a number of interesting images already and curbing a small addition to the excitement of making such objects, I elected to not proceed further with refining my metalwork creations.

• 

• 

• 

• 

• 

With that said, and knowing they were not perfect, and with lots of odd little discrepancies they nevertheless look pretty good. Convincing enough.

• 

• 

• 

• 

• 

• 

Traditional Chinese Paintings

Now this was an area I feel I could explore in more depth with a bit more time. Fascinating results. Especially when you offer an existing image to convey style. But again, time and money are limited.

Climbing back out of the rabbit hole

As I am sure you have now guessed, this is a deep rabbit hole, in which one could easily spend hours upon hours (something I can attest to) designing new prompts and conjuring up novel fantastical images. And I have barely even scratched the surface considering the complexity of training models, weighted prompts and other more advanced methods.

Highly addictive.

The anticipation, the creative aspect, the satisfaction of an realistic image… All very addictive.

And it is for this reason that the roads ends for me here.

But the technology will continue to evolve and offer finer, more granular control thus allowing for even more realistic images. Soon one will be able to create the same image but depicted from multiple angles. Additional refining of the data set will be possible. And smaller and smaller details will continue to be tweaked and adjusted. As it evolves it will bring up a range of interesting questions to ponder. Questions such as, “would it be possible to use the AI to create new, true-to-form styles “antiques” which could be produced as high end reproductions?” The backstory would be fascinating. Or could criminals use it to imagine a new painting in the style of a particular artist, which can then be handed off to a skilled forger, who would then create the physical painting? How about trainers and authenticator using them as “test images” to assess one’s skill in recognizing if something is off? Scammers could create unique items that look credible with images of the same item from multiple angles. These amazing finds could then be sold (but never shipped) to unsuspecting buyers.

I guess we will have to wait to see what the future holds.

Hopefully we will avoid the Hollywood movie version of a sentient AI destroying civilization. 😉

I leave you with a few Android dreams.

All the images you see here, with the exception of screenshots and training images, were generated by an AI.