The problem: 78 cards that need to look like one deck
Anyone who has tried to generate a full tarot deck with image models knows the pain: every card comes back in a different style. One card is photorealistic, the next is flat vector art, and the third looks like it was drawn by three different artists. A tarot deck is fundamentally a design system — 78 individual illustrations that must share a visual grammar: linework, palette, framing, typography, and composition rules.
The Seven of Wands shown here is a good example of what "locked style" looks like in practice, so let's break down how a card like this gets produced and what engineering decisions sit behind it.
Anatomy of the card
Look at the structure before the art. The image is built in layers:
- A decorative frame — an ornamental border of gold vine and thorn motifs on a dark ground, with star-shaped medallions at the corners.
- A header plate with the Roman numeral VII, and a footer plate with the card title in a display serif face. These plates are part of the deck's shared chrome, identical across cards.
- The central illustration, in a detailed comic/fantasy-ink style: a defender on high ground wielding a staff against multiple attackers thrusting wands upward from below, set against a sunset sky with radiating light.
That layered structure is deliberate. If you bake the frame, title, and numeral into every generation, you lose control. The reliable approach is to generate the illustration only, then composite the chrome programmatically.
The pipeline
A practical production flow looks like this:
Style anchoring. Pick one strong reference card and reuse its style descriptors verbatim in every prompt: line weight, shading style (heavy ink outlines with cross-hatched color), palette (warm ochre and rust tones), and mood keywords.
Composition constraints per card. Traditional Rider-Waite-Smith iconography gives you the scene: for the Seven of Wands, a figure on elevated terrain defending against six wands from below. Encoding the traditional meaning in the prompt keeps the deck readable for actual tarot users — this is domain knowledge, not decoration.
Negative-space planning. The footer plate needs a clean band for the title. Prompting for a bordered composition with dark margins gives you predictable space for post-processing overlays.
Post-composite. The numeral and title typography are added in code — SVG or canvas rendering over the generated art:
const svg = ` <svg viewBox="0 0 1024 1024"> <image href="card-art.png" width="1024" height="1024"/> <text x="512" y="64" text-anchor="middle" class="numeral">VII</text> <text x="512" y="975" text-anchor="middle" class="title">The Seven of Wands</text> </svg>`; This guarantees that all 78 cards share pixel-identical typography, kerning, and placement — something no image model will give you consistently.
- Batch QA. Generate contact sheets of all cards side by side. Style drift is far easier to spot in a grid than one card at a time. Regenerate outliers with tightened style tokens.
Results and trade-offs
The layered approach pays off in the details you can see in this card: the frame ornament is crisp and symmetrical, the numeral sits exactly centered, and the title type is legible and consistent — because none of that was left to the model. The model does what it's good at (expressive narrative illustration, dramatic lighting, dynamic figure posing), and code does what it's good at (repeatability).
The trade-offs are real, though:
- Text in-model is unreliable. Never let the model draw your card titles; baked-in text comes out mangled at scale.
- Style drift compounds over long batches. Re-anchor prompts against the reference card every few generations rather than iterating from the previous output.
- Human review stays in the loop. Iconographic accuracy (which way the wands point, how many there are) needs a person who knows the deck's tradition checking each card.
The end result is a deck where card VII looks like it belongs next to cards VI and VIII, and the whole set can be published as a coherent web experience — which is exactly what we've done. You can view the full deck at the project here: https://bevespera.top
If you want to follow along or ask questions about the pipeline, there's a WhatsApp group where progress and discussion are shared: https://chat.whatsapp.com/KVjWIzwkqaYLEOqbRze4R1
AI-generated tarot card for the Seven of Wands: a layered composition with an ornamental gold vine-and-thorn border frame, corner star medallions, a header plate with the Roman numeral VII, and a foot



