Tarot Deck in a Weekend: 78 Cards, One Locked Character

I spent three months trying to make a personal tarot deck with AI before I admitted the workflow was wrong. Card by card, the character would drift. The Fool on card one looked different from the Fool on card twenty-two when I had to redo it. The Page of Cups looked like a stranger to the Page of Wands. By the time I had thirty cards I had three or four versions of the same protagonist and a deck that did not read as one universe.

The problem was not the model. The problem was that I was building the deck card by card instead of building it from a locked anchor. Once I switched to a lock-once, generate-everything workflow, I shipped a coherent 78-card deck in a single weekend. The principles transfer to oracle decks, tarot variants, and any other large card series where consistency matters more than artistic ambition.

This is the workflow. Two days. Seventy-eight cards. One character that holds across the whole deck.

Quick Answer: For an AI tarot deck, lock the character with IPAdapter plus a small style LoRA. Treat the Major Arcana as the style source for the entire deck. Generate the Minors as variable compositions within the suit identity frame anchored by their court card. Plan for a quality pass on the twelve cards that historically drift. Two-day production schedule is realistic for a solo creator with this workflow.

Why Most Personal Decks Stall at Card Twelve

The standard story for AI tarot deck attempts. You start strong with The Fool, The Magician, The High Priestess. Cards one through ten are usually beautiful because you are excited and you spent real time on each. Around card twelve, fatigue hits. You start cutting corners on prompt detail. Two or three cards in a row come out lower quality. You lose momentum. The deck sits unfinished for three months.

Even if you push through, the consistency starts to degrade. Card twelve looks marginally less like card one than card eleven did. The drift is gradual but compounding. By card thirty you are visibly working with a different protagonist than you started with.

Then there are the Minor Arcana. Fifty-six cards split across four suits, each with ten numbered cards and four court cards. The numbered cards do not have established imagery in many decks. You are inventing compositions. The court cards have to feel like nobility but also like the same person across all four suits. The volume is daunting. The creative load per card is heavier than the Majors.

Most personal decks die here. The creator has spent their best energy on the first twelve cards, the workflow is unsustainable, and the back half of the deck either never gets made or gets made at a visibly lower standard.

The fix is to invert the workflow. Lock the character once. Set the style once. Then generate everything as variable compositions inside that locked frame. The creative energy goes into the lock and the style anchor, not into each individual card.

The Style Anchor, One Look That Carries the Whole Deck

The Major Arcana is your style anchor. Twenty-two cards. Generate these first, generate them carefully, and accept that they will set the visual contract for the rest of the deck.

Why Majors first. The Major Arcana cards are the most iconic. Readers recognize The Fool, The Sun, The Moon, The Tower at a glance. They have established imagery to draw from. The mental templates are strong. This makes the Majors the easiest cards to generate with confidence because both you and the model know what they should look like.

When you generate the Majors first, you are doing two things at once. You are building the deck and you are calibrating the style. Each Major card teaches the workflow how the deck should look. By card twenty-two you have a fully developed visual vocabulary that you can carry into the Minors.

The discipline. Do not move on to the Minors until all twenty-two Majors are in the deck and approved. If one Major still does not look right, fix it before starting on the Minors. The Minors will inherit the style of the Majors. Inconsistent Majors produce a chaotic deck.

A practical structure. Spend Saturday morning on the first twelve Majors. Take a break. Spend Saturday afternoon on the remaining ten Majors. By Saturday evening you have your full Major Arcana locked. Day two is the Minors.

The Major Arcana as Style Source for the Suits

This is the trick that took me three attempts to figure out. The Major Arcana does not just set the style. It sets the style source for each suit.

Each tarot suit has a thematic element. Cups are water, emotion, and intuition. Swords are air, intellect, and conflict. Pentacles are earth, material, and abundance. Wands are fire, passion, and creativity. Each suit needs a visual identity that ties to its theme.

The Major card I anchor each suit to is selected by theme. Cups inherits style from The Moon. Swords inherits from Justice. Pentacles inherits from The Empress. Wands inherits from The Sun. The four anchor Majors are chosen for thematic resonance with their suits.

What "anchored" means in practice. When I generate the Page, Knight, Queen, and King of Cups, I use The Moon as a style reference image alongside the character reference. The model receives both the character lock and the visual context of The Moon. The resulting court cards feel like they belong in the same universe as The Moon while being recognizably about Cups.

The same pattern applies to numbered cards. Three of Cups inherits The Moon as the style anchor. Five of Pentacles inherits The Empress. The full deck ends up with a cohesive visual language because every card traces back through its suit anchor to a Major card you already generated and approved.

Designing the Four Suit Identity Frames

After the Majors are done, before generating any Minors, design four identity frames. One per suit. The identity frame is a short style and composition guide that defines how every card in that suit will look.

For Cups, my identity frame is something like: "soft moonlit palette with deep blues and silvers, water elements present in every composition, the protagonist in flowing garments, dreamlike soft-focus quality." Every Cups card from Ace through King uses this frame.

For Swords, my frame is: "sharp linework against cold gray skies, wind and sharp objects present in every composition, the protagonist in structured clothing, high-contrast lighting." Every Swords card uses this.

The identity frames live in my prompt template as the fixed-clause section for that suit. Generating a Cups card means assembling the protagonist identity clause plus the Cups identity frame plus the specific card's variable clause. Generating a Swords card swaps the Cups frame for the Swords frame and keeps everything else.

This is the single biggest accelerator in the workflow. Once the four frames are written, generating a Minor card is just selecting the right frame and writing the card-specific variable clause. The creative load drops to maybe a quarter of what it was when each card needed its own full prompt.

Number Cards, Variable Composition With Fixed Style

Numbered Minor Arcana cards are where the deck either holds together or falls apart. There are forty of them, ten per suit. They do not have the established imagery of the Majors or the inherent nobility of the courts. They are the toughest creative load in the deck.

The workflow for numbers is variable composition within a fixed style. The composition changes per card to express the card's meaning. The style does not change at all within a suit.

For each numbered card, the variable clause specifies:

• How many subjects are in the frame (the number itself)
• The action or stillness the card depicts
• The emotional register

Two of Cups is two figures in connection. Five of Cups is one figure mourning over three spilled cups while two stand behind. Eight of Swords is one figure bound and surrounded by swords. The variable clause expresses these specifically.

The number on the card has to be readable. Either there are literally N objects in the frame (N cups, N swords, N pentacles, N wands) or there is a composition that implies N. Most decks include the literal count because it is easier for readers. I do this in my own decks because it serves the reader, not because it is artistically more interesting.

Practical note. Numbered Minor cards generate faster than Majors or courts because the variable clause is mostly composition logic. I can generate ten numbered cards in a suit in maybe ninety minutes including the regeneration loop for the ones that miss. Four suits times ninety minutes is six hours of generation time for all forty numbered cards. That fits in a day.

Court Cards and the Identity-Sharing Problem

The four court cards per suit are sixteen cards total. They have to feel like nobility, royalty, and figures of authority. They also have to feel like distinct individuals within each suit. And they have to relate to the protagonist character lock without being literally the protagonist.

Here is where most decks struggle. If you use IPAdapter on the protagonist for every court card, you end up with sixteen cards of the same person wearing different costumes. That is not what a court is. A court is sixteen related characters who share a universe but are distinct people.

The fix is a related-character pattern. Use the protagonist as the loose visual reference but with explicit instructions to vary age, gender presentation, and specific features. The Page of Cups is a younger figure, perhaps adolescent. The Queen of Cups is an older figure with maternal energy. The King of Wands is a powerful figure with leonine bearing. These are described in the variable clause and the IPAdapter weight is reduced to maybe 0.4 to 0.5 so the protagonist influences but does not dominate the courts.

The other trick. Generate the four Kings first across all suits. They are the strongest characters and they set the bar for the courts. Then generate the four Queens. Then Knights. Then Pages. This top-down order means the courts feel like a hierarchy because they were generated as one. Going Page-Knight-Queen-King per suit produces less cohesion because you are processing the courts in a different order than they exist in the deck.

Card Borders, Titles, and Print-Ready Resolution

The generated illustration is the inside of the card. The full card needs a border, a title, suit symbol, and number. This happens in layout after generation, not in the prompt.

For borders, design one master border that holds the entire deck. Suit symbol placement, title font, number placement. The master border applies to all 78 cards. Variation per suit comes from a subtle color shift or symbol style. Cups borders might be silver-blue, Swords might be steel-gray, Pentacles might be earth-gold, Wands might be flame-red. The structure stays identical.

Title font is critical. Pick a font that has tarot-deck personality without being illegible. Avoid display fonts that fight the illustration. Avoid generic system fonts that read as cheap. A good serif with deck history works well. The Universal Tarot, the Rider-Waite, and most published decks use serif fonts in the title position.

Resolution. Tarot cards are typically printed at 70mm by 121mm or thereabouts. At 300 DPI, that is roughly 825x1430 pixels per card. Most AI image generators produce more than enough resolution for this. The bigger concern is bleed. Add 3mm bleed on all sides for print-on-demand services. The illustration needs to extend into the bleed area without important elements being clipped at the trim line.

For print-on-demand tarot deck services, MakePlayingCards, MPC, and DriveThruCards are the standard options. Each has slightly different specs. Confirm bleed and trim sizes before exporting.

Quality Pass, The Twelve Cards Worth Reviewing

After all 78 cards are generated, there is a mandatory quality pass on the twelve cards that historically drift. These are the cards where my decks have most often missed and where I now always check before considering the deck done.

The twelve cards to check carefully:

• The Fool (often the first card generated, identity loosest)
• Death (model bias toward generic skull imagery)
• The Devil (model bias toward generic horror imagery)
• The Tower (composition complexity tends to fail)
• The High Priestess (gendered features sometimes drift)
• The Hermit (older-character anchor sometimes misses)
• Eight of Cups (cup-counting often wrong)
• Nine of Swords (composition tends to fail on the nightmare imagery)
• Ten of Wands (object-count and burden composition often fails)
• All four court Pages (younger-figure anchor sometimes misses)

If any of these miss on the initial batch, regenerate them with adjusted parameters. Higher IPAdapter weight on the character-sensitive cards. Tighter variable clauses on the composition-sensitive cards. Different seeds on the cards that need creative variation.

The quality pass usually adds two to four hours to the deck production timeline. It is worth every minute. A deck where ten cards are clearly worse than the other 68 is a deck that nobody will use seriously.

Two-Day Production Schedule You Can Actually Hit

Here is the actual schedule that fits into a normal weekend.

Day One, Saturday.

Morning, four hours. Build the protagonist reference set (three images). Write the prompt template with fixed identity clause and fixed style clause. Generate the first twelve Majors. Approve or regenerate as needed. End of morning, twelve Majors locked.

Afternoon, four hours. Generate the remaining ten Majors. Approve and regenerate. Design the four suit identity frames. End of afternoon, twenty-two Majors locked and four suit frames written.

Evening, two hours. Generate the sixteen court cards across all four suits, going Kings, Queens, Knights, Pages. End of day one, all 38 character-driven cards are done.

Day Two, Sunday.

Morning, four hours. Generate all forty numbered Minor cards. Ten per suit at maybe ninety minutes per suit. End of morning, all 78 cards exist as generated illustrations.

Afternoon, three hours. Run the quality pass on the twelve high-risk cards. Regenerate as needed. Lock in the final 78.

Evening, two hours. Layout pass. Apply master border. Add titles, numbers, suit symbols. Export print-ready files.

Total time invested. Roughly nineteen hours across two days. That is achievable for a solo creator who blocks the weekend for the project.

If you cannot block a full weekend, the workflow stretches across four shorter sessions of about five hours each. Day one for Majors and suit frames. Day two for courts. Day three for numbered cards. Day four for quality pass and layout. Same nineteen hours, distributed.

Running the Whole Deck Through One Apatero AI Workflow

Apatero AI has a deck workflow that automates the prompt-template assembly, the suit-frame routing, and the character lock across all 78 cards. Inputs are a protagonist description, a style anchor, and a 78-card variable-clause list. Outputs are 78 print-ready illustrations.

The platform handles the order automatically. Majors first, courts by hierarchy, numbered Minors by suit. The character lock applies to every card. The suit frames apply correctly per card. The quality pass flags the twelve high-risk cards for manual review.

Full disclosure, I help build Apatero AI. The deck workflow exists because I shipped a personal deck on this manual pipeline and wanted to make the workflow available to anyone who wanted to ship their own. The manual workflow described in this article works fine. The hosted workflow saves about half the orchestration time.

For deeper background, How to Lock a Character Across 50 Images With Apatero covers the character-lock fundamentals at smaller volume, Character Sheet From One Reference covers the reference-set construction, and Lighting Prompts That Hold Across an Image Pack covers the lighting-consistency techniques that matter for the per-suit visual cohesion described above.

FAQ

How many reference images do I need for the protagonist?

Three at minimum. Front view, three-quarter view, side view. For complex protagonists with distinctive back-of-head features, add a back view. The reference set takes thirty minutes to build and saves hours on the 78-card generation.

Should I generate the Majors all in one session?

Yes, ideally. The stylistic calibration that happens during Major generation is most consistent when done in one continuous session. Splitting the Majors across sessions can introduce subtle style drift that affects the rest of the deck.

Can I use the same character lock for multiple decks?

Yes. If you build a strong reference set for a protagonist, you can reuse it for tarot deck variants, oracle decks, lenormand spreads, or even a Marseille-style deck with the same character. The lock is reusable.

What if I want a non-human protagonist?

The workflow accommodates this but with caveats. Non-human protagonists benefit from a small LoRA in addition to IPAdapter because the model has weaker identity representation for non-human categories. Add about four hours to your timeline for LoRA training before starting the deck.

Do I need to be able to draw to make an AI tarot deck?

No. The entire workflow is prompt-driven plus reference-based. Drawing skill is helpful for designing the master border in layout but the illustrations themselves do not require traditional art skills.

How do I avoid the deck looking generic?

The protagonist character lock is the single biggest differentiator. A generic tarot deck has stock-looking figures. A deck with a strongly-defined protagonist that appears across all 78 cards reads as distinctive. The style anchor is the second differentiator. A clearly-articulated style choice in the Majors carries through the entire deck.

Can I sell a tarot deck made this way?

Yes, but the legal landscape for AI-generated commercial art is evolving. Best current practice is to make substantial creative contributions through prompting, curation, and layout, to use AI as a tool rather than as the sole author, and to consult an attorney for commercial sales. Many AI-illustrated decks are selling successfully through Kickstarter, Etsy, and direct-to-consumer channels.

What models work best for tarot illustration?

In my testing, Flux for photorealistic and semi-realistic decks, SDXL for stylized illustrative decks, and the major closed models for variety. The character lock workflow is model-agnostic. The same IPAdapter-plus-reference approach works across all of them with parameter tuning.

How do I handle the Justice card specifically?

Justice has heavy symbolic content (scales, sword, blindfold). The variable clause needs to include all three symbols explicitly. The model sometimes drops one of them on the first pass. Always regenerate Justice if any symbol is missing.

What if I want a deck that includes the Lovers in a non-traditional way?

Variable clauses are flexible. Generate the Lovers however your deck's interpretation requires. The character lock is on the protagonist, not on the specific symbolic configuration. Custom interpretations are part of what makes a personal deck personal.

Wrapping Up

The 78-card deck is achievable in a weekend. Lock the character once. Build the suit frames after the Majors. Generate top-down through the courts. Run the quality pass on the twelve drift-prone cards. Layout and export.

If you do not want to manage the workflow manually, the Apatero AI deck tab compresses the pipeline into one job. For external references, the VDraw tarot generator covers single-card generation, the Pixelcut deck workflow shows an alternative production model, and the card design specs at MakePlayingCards cover the print-ready file requirements for offset deck production.

The takeaway. Decks fail when you build them card by card. Decks succeed when you lock the character first and treat every card as a variable within the lock.