List of tools used
OpenAI API - https://platform.openai.com/
GPT4 / ChatGPT
Dall-E 3
Background: What is Euchre?
Euchre (YOOK-err) is a card game from the United States Midwest. It uses an abbreviated deck, has four players with two teams of two, and play involves winning (“taking”) tricks where each player must play a card from their hand, with the winning cards based on a trump suit selected at the beginning of the rounds.
That is to say, I’ve played a lot of euchre, while drinking, with my family (specifically my mom). As generative AI booms, what better way to take advantage of it than to make some custom playing cards as a gift?
The goal:
Use generative AI to create custom playing cards. (Success)
Generative AI should also design the playing cards, using themes that are special to my family. (Success)
Run the images through Photoshop to prepare for printing. (Success)
Send them off in time to get them for 2023 Christmas gifts. (Fail)
AI Is Bad at Playing Cards: Using Midjourney to Prototype
My initial attempts at art used Midjourney via the Discord app. This generates four images per prompt and allows for fairly rapid iteration. But let’s just say the results weren’t what I wanted. However, I did narrow down on the art style I wanted: Black backgrounds, Red (Hearts and Diamonds), Gold (Spades and Clubs), art deco, and intricate.
My favorite part of these images is how inconsistent the playing card symbol shapes are. For something that has such a clear standard, it’s fascinating that Midjourney cant keep them straight (we’ll see that Dall-E is only a little better). I suspect it’s because the training data doesn’t have a lot of specific cards (which cards are which) to work from - the classic Ace of Spades seems to have a little too much influence over everything else.
Midjourney wasn’t giving me the results I wanted but at this time (November 2023), Dall-E 3 had just been released by OpenAI and by all accounts was ahead of Midjourney in many ways. Plus, my partner happens to already have an OpenAI subscription which gives us access to the API, not just the prompt window. After about Midjourney three attempts to make different King cards (which I thought would be some of the easiest), poor results, and a daunting 24 cards to complete, I was more than ready to have an AI write the prompts for me too.
Theming the Cards: Prompting in ChatGPT
To start, I used ChatGPT4 to write the prompts.
“I am making a custom deck of euchre cards for my family. Can you help me design some cards based on some family traits? The design style I have in mind is a black background with red or gold foil accents. The design should be art deco, with stylized decorations and a limited color palette, red/gold or black only.
Some of the themes I want to include in these playing cards are: New Mexico (mountains, the Zia symbol, sand dunes), mountain biking, bicycling, parrots (cockatiels), art (colored pencils and acrylic painting), mecha robots (Gundams), airplanes, astronomy and stargazing, big city skyscrapers (Seattle and New York City), crystals (amethyst).
What are some themes we could use for each suit that would bring these ideas together?
Can we integrate astronomy in each theme?”
After some relatively minor iteration, we got the following themes. These didn’t fully stick through the final selections, but was a great starting point.
Hearts: Nature, Art, and the Cosmos
New Mexico landscapes (mountains, Zia symbol, sand dunes) with a night sky backdrop featuring stars.
Art supplies (colored pencils, acrylic paints) arranged in a pattern resembling constellations.
Crystals (amethyst) with a starry gleam or designed like celestial bodies.
Diamonds: Adventure, Exploration, and Celestial Navigation
Mountain biking and bicycling under a starlit sky.
Airplanes flying high among stars, perhaps with a compass or celestial navigation tools.
Astronomy elements like telescopes, stars, and planets integrated into the biking and flying scenes.
Clubs: Urban Life, Technology, and Cosmic Influence
Skyscrapers (Seattle and New York City) with a starry sky above, highlighting the contrast between urban life and the universe.
Mecha robots (Gundams) designed with celestial motifs or star patterns.
Technological elements infused with astronomy (e.g., buildings with observatory domes).
Spades: Natural Connections and the Universe
Parrots (cockatiels) with feathers patterned like constellations or galaxies.
Nature scenes with prominent starry backgrounds, emphasizing a connection between earth and sky.
A mix of organic shapes and celestial bodies, like trees reaching towards the stars.
Robot, Do the Boring Parts: Using the API and Google Colab to generate full prompts and images
system_prompt = """You make prompts for dall-e 3 for custom playing card designs. You provide the outputs in json in this format:{"number": 5,"suit": "Hearts","prompt": "..."}Each card should be in a simple, two-tone art deco style. The heart and diamond cards should be red and white on a black background. The club and spade cards should be gold and white on a black background. The central art should be frame by stylized art deco elements.For each card, you can create a central motif that embodies the theme of the suit, surrounded by decorative art deco elements. The use of black backgrounds will make the red and gold foil accents stand out, giving the cards an elegant and timeless feel.Remember, the key to art deco design is symmetry, bold geometric shapes, and lavish ornamentation. This style will beautifully encapsulate the various elements of your family's interests and traits."""example_in = """9 of Spades: A mountain biker on a trail at night, guided by the light of the Spica star."""example_out = """{"number": 9,"suit": "Spades","prompt": "A playing card design in a two-tone art deco style, featuring the 9 of Spades for a custom Euchre deck of cards. The design is in black and gold, depicting a mountain biker on a trail at night, guided by the twinkling light of the Spica star. The background is a rich black, emulating the night sky, adorned with gold lines that form decorative art deco elements and celestial bodies. The biker, depicted in an art deco fashion, is silhouetted against a backdrop of stylized stars and planets, with spade-shaped constellations prominently featured in the sky. This card design uses a taller aspect ratio to encapsulate the adventurous spirit of mountain biking intertwined with the mystery and beauty of astronomy."}"""
^^^ The general prompt for our little API program. It consumed a single line description, and was directed to generate a full detailed description. It also needed to track the number and suit so that we could generate consistent unique file names for the output images and know what was what.
The single line descriptions (“example_in”) came from ChatGPT, pasted into the notebook directly, with some human editing. We ran at least 3 full sets of prompts and images, separating them within the code to preserve the record and allow us to switch the run by modifying a version number.
It did take some iteration to get consistent JSON outputs. Sometimes AI is too helpful…
Both the “Gundam” and “Cockatiel” prompts sometimes ran into OpenAI content filters. “Gundam” usually went through after a try or two, but several prompts I manually replaced “cockatiel” with “Nymphicus hollandicus” to get my birds to generate, which worked very well.
A key feature was to check for an existing file before submitting the prompt. Consistent file names were key, and this allowed us to easily rerun the code when a few prompts failed without overwriting successful past attempts, or manually editing which ones could run.
All output full prompts were also recorded to be able to reference against the output images.
Link to download the .ipynb file at the end of the post. This file was edited to remove the API key and the personalized prompt. I cannot provide support for making this work for you - Sorry! I am not familiar with this tool and most of the work was done by my partner.
Best of Dall-E 3 Outputs (No cropping or editing)
Card Curation: I love bad AI art.
I want to emphasize: AI still has a lot of work to do in terms of understanding and portraying the real world. The art style here is fantastical, which forgives many sins, but there are obvious errors that a human artist would have never made. The more you look, the more you’ll find them.
I curated the art based on what I thought were the prettiest and most fun designs for the most part - and did not reroll for perfect art. In fact, AI art errors were part of the goal - I wanted a deck of cards that as you played, you would look at a card in your hand and find some stupid feature from the AI. I wanted the conversation about the art quality and it’s failures, not just a pretty deck of cards. This is not a project that’s just about how to use AI to make unique customized products, but also a project to enjoy the hilarity of the current state of AI absurdity.
We generated less than 200 designs in all using Dall-E, with three full generations and some re-rolls for specific card failures. This took about one evening’s effort - from starting code to having all the art generated and ready to choose from.
Best of the Worst Image Gallery (Both Midjourney and Dall-E)
Photoshop Modifications: Making a functional deck
You might notice something from the images above: These AI generated images are terrible playing cards. The art is cool (usually) but they vary wildly from image to image. The corner numerals and pips are the most wild - go back and pay attention to the corners and try to guess which card was the actual prompt. Spoiler: You can’t, the AI could not generate card identities in the art consistently. Clubs were particularly difficult - there are many spade shapes across the Club cards I generated. The aspect ratios allowed by the generators are inconsistent, and forget about things like bleed and foil layers. Trying to play with a deck of cards with chaotic iconography would be fun - but wasn’t the game I was going for.
This meant every card needed to be changed and refined to have a unified look. I opted to add a solid background behind the numerals and pips to make them stand out. I also added borders, to create that important bleed and safe zone, and a consistent size. At the end of the post you can find a download button for the files to save you some of this effort.
Not Doing That Again: Foil Printing
Long story short: the back and forth to get the files correct for the printer took weeks. I tried to ask for help, but got no useful information. In fact, what I got was misleading and contributed to the first of multiple re-dos. Foil printing requires a proof in addition to the base image. And even the base image needed a few adjustments - they were concerned my “safe zone” was too small, which could lead to things getting cut off given the margin of error during printing and cutting, leading me to shrink everything to add a larger margin. I made about 4 rounds of updates, which required making new versions of 48+ files each time. These recommendations and updates were all made in good faith and wanting to ensure I was happy with the final result - but it was painful. Literally - my carpel tunnel was very angry at the repetitive changes.
My cards arrived the day after I flew across the country for Christmas plans - so I had to mail all of the cards as post-Christmas gifts sometime in January. Very disappointing, but fine I guess in the grand scheme of things.
Given this experience, I won’t be naming the company I used. You’ll just have to guess!
Extras
The back I designed myself, with only a few accents coming from some of the image generations. This was too fun to let a robot do it, and I had a clear visual I wanted to create. As should be obvious by now, my dad does a lot of mountain biking (the tire), and my family is based in New Mexico (the Zia symbol). My mom’s art frequently features citrus fruits, which is why there’s an inner orange.
I made a rules card - It was rough trying to explain Euchre in such a small format. I got some help from friends for proofreading and suggestions (perks of having board game nerd friends). I’m not the most happy with the result, but it works if you already know the basic rules and just need a reference for the complicated bits and scores.
I made score cards. When playing with a normal deck, the tradition in my family is to grab all the 5 cards and use two per team to track scores. This is apparently controversial, so I can only say this is what my family has always done. Since these cards would be missing in my deck, I made 5’s for each suit to use for scoring. I originally made a set of custom pips, but realized they would only work at larger print sizes, so I reused them on these score cards.
I purchased a large pack of hard plastic cases from Amazon, sized for decks of cards for each gift. I looked into printing custom boxes as well, but they would have doubled the cost of this already too expensive project.
Every deck I’ve given away had a note stating the art was generated with AI and a brief description of the process - I wanted something that in 5-10 years someone could find one of the cases and understand what had happened. (A minor retrospective: I printed stickers and these should have gone directly onto the cases so they could not be lost or thrown away)
Regrets, Retrospective Changes
I would not do foil printing, that was too much work for the output. It doubled the amount of repetitive files I needed to create. I would not use an international company but stick with someone in my hemisphere for timezone reasons. Not to mention - foil printing made this way more expensive! (Don’t ask how much I money and time I spent on this project, I don’t want to talk about it)
I would plan a full set of cards, not the abbreviated deck. It would have been only slightly more difficult to get a full deck of cards that could be used for any game, since so much was automated (the art at least!). In addition, it seems that many custom card deck printing options are designed for 52 cards + 2 jokers + 1 extra - something about 55 cards fitting on a standard sized sheet. I might have been able to get a cheaper per deck price for a larger deck.
I would have done a few more generations. Some cards (Ace of Diamonds, looking at you) I didn’t really care for the end result, but I was a little too exhausted to keep going at the time.
Plan ahead and make the Photoshop easier. The Photoshop step was by far the most tedious. I would have spent more time initially with the layer and file set up to make each card easier. There are also some macros or automation that could be done and that might’ve been worth learning and doing. After the third revamp, I had a system, so it would have been nice to have that in the first place.
I would have used Photoshop to unify the colors. While the final outcome had a lot of consistency, it’s not perfect (the Ace of Hearts bothers me the most). An overlay or two could have made the final cards a little more cohesive with one another without a lot of effort.
I would get a more basic card paper option. I got a fancy “soft touch” coating on the cards. It was a little more expensive, and it ends up it’s not very good. It makes the cards a little velvety - which is nice to touch but it makes shuffling and actually playing with the cards harder. They stick together a little bit. It’s just not practical - so annoying to have paid extra for a worse outcome.
I would make physical suit tokens. Euchre has a trump suit that needs to be remembered. A small set of 3D printed tokens would make this much easier when playing. This is still on my to do list - maybe next year’s holiday gift.
Final Outcome
Overall, I’m proud of the result. I love most of the art, and it’s fun to find AI errors and weirdness while playing. I carry around one of the decks (of which I have too many extras) with me to show off.
My mom lives across the country, but we were able to meet and play together (we beat her and my dad handily, whoops!), and it was really fun to be able to take a hobby she liked and make a big thing of it. Unfortunately, it plays only okay - as mentioned, the card texture is a little sticky. The colors work for the most part to differentiate the suits, but there is so much noise on each card (the art is busy by design), it still adds mental processing time to figure out if you are holding a clubs or a spades. Most playing card designs are simple for a reason.
Would I do it again? Probably not, at least not until I forget more about the annoying parts. But it was an interesting enough project that I’m glad I went through with all of it.
Thank you!
I’m sharing this process to help anyone else who is looking to use AI for custom art projects. I really love the ability for generative AI to give someone who does not have the skill or time to create art themselves, or money to pay for an artist, to create beautiful art of a wide variety. My vision for generative AI is a world of more beautiful things because we can make things we could never do otherwise. As the genie is out of the bottle, we’re still learning how things will unfold.
There are many valid controversies to using generative AI. I am skeptical of using pure AI art generation - like what I made - in a for-profit venture. I am not selling these cards; they were gifts. We will each have to make our own decisions on how to use this technology, and I support government intervention and regulation that provides safeguards for original creators.
I hope you found this an interesting and/or useful read. I am not a blogger, but I’m a big believer in documenting successes and failures to help others in their own journeys.