Feeding a portrait to an image model and asking for "a fantasy version" gets you a stylised portrait. Asking for a collectible card gets you something structurally different: a framed artifact with a border, a name plate, a rarity marker and a posed subject inside the art window. The second is far more useful, and it comes from one change in how the prompt is written. Here is the structure, the reference-photo rules that decide whether it works, and the limits worth knowing before you publish anything.
Describe the artifact, not the person
The key move is naming the object you want rather than the image you want. "A portrait in fantasy style" describes a picture. "A collectible card game card depicting this character as a legendary unit" describes a physical thing the model has seen thousands of examples of — and card layout is a genre with strong conventions. Ask for the genre and you inherit the frame, the ornamental border, the art window, the name banner and the general composition for free, without listing any of them.
This is why the shortest version of this prompt is often the best one:
Reference image + prompt: A collectible card game card, in the style of Blizzard's Hearthstone, depicting this character as a legendary card. Redraw the character or subject in whichever illustration style suits them best, in a dynamic pose, so that the result is appealing.
Two details in there are doing most of the work. "Legendary" is not decoration — in card games it is a rarity tier with its own visual language, so it pulls in the more ornate frame treatment. And "redraw in whichever style suits them best" deliberately hands the style decision to the model, which tends to produce a more coherent illustration than forcing a style that fights the subject.
A reusable template
When you want control rather than a lucky result, expand the same idea into named slots:
A collectible trading card, portrait orientation, with an ornate metallic frame and a clear art window. Inside the art window, the person from the attached photo, redrawn as a painted fantasy illustration in a dynamic three-quarter pose, holding their gaze toward the viewer. Deep blue and gold palette, dramatic rim lighting, painterly brushwork. A name banner across the lower third reads "THE CARTOGRAPHER". A small rarity gem sits at the top of the frame. Keep the face recognisable from the reference. No other text anywhere on the card.
The slots, and what each one controls:
| Slot | Controls | Typical failure without it |
|---|---|---|
| Card format | Orientation, frame material, art window | A plain illustration with no card structure |
| Subject source | Which face gets used | A generic person unrelated to your photo |
| Art treatment | Painted, inked, cel-shaded, 3D render | Style drifts between runs |
| Pose | Angle and energy of the figure | Stiff, front-facing, passport-photo posture |
| Palette and light | Colour scheme, lighting direction | Muddy contrast between art and frame |
| Name plate | The one string you want rendered | Invented gibberish text in the banner |
| Exclusions | Everything you do not want drawn | Stray stat numbers, watermarks, extra captions |
The reference photo decides most of it
Model choice matters less than input quality here. A good reference is:
- Sharp and well lit, with the face unobstructed. Soft or backlit inputs produce a subject the model has to invent, and invention is where likeness goes.
- Head and shoulders or half body. Full-body shots leave the face too small to carry identity through a redraw.
- Neutral in expression. Strong expressions fight the pose you are asking for.
- Free of heavy filters. Beauty filters and aggressive grading get baked into the stylisation as artifacts.
- One person only. Multiple faces in the reference reliably produce a blended one.
If likeness is drifting, fix the reference before you touch the prompt. Adding "keep the face recognisable" helps at the margin; a better photo helps far more.
Iterating
- Run the short version first. If the card structure appears at all, the model understands the genre and the rest is tuning.
- Lock the layout before the style. Frame, art window and banner position are what make it read as a card; palette can change later without breaking anything.
- Change one slot per run. With this many instructions, a shotgun edit leaves you unable to tell what helped.
- Handle the name plate last, and keep the string short — a couple of words render cleanly where a sentence does not.
- If text keeps coming out wrong, generate the card with an empty banner and set the type yourself afterwards. This is nearly always faster than fighting for a perfect render.
Which models this works on
You need image conditioning — the ability to pass a reference photo alongside the text — and an encoder that can hold several distinct instructions apart. Hosted models from the major image labs handle both well. On the open-weight side, the image-editing and reference-conditioned variants of the current model families can do it locally, at the cost of setup and VRAM.
One behaviour differs meaningfully between models: how literally they preserve a face. Some prioritise likeness and under-stylise; others commit to the art style and lose the person. If a model gives you a beautiful card that is not recognisably your subject, that is usually a property of the model, not a prompt you can rescue.
Limits worth knowing
- Named-property styles. Invoking a specific commercial game's card look is effective precisely because that look is distinctive and protected. Fine for a personal avatar or a gift; a poor idea for anything you sell or use in marketing. Describing the visual conventions generically — ornate frame, rarity gem, name banner — gets you most of the way without borrowing a trademark.
- Likeness. Someone else's face needs their agreement, whatever the prompt produces.
- Card text. Stat blocks and rules text render badly. Ask for a clean card and add real text in a design tool if you need it.
- Series consistency. A set of cards for different people will not share a frame exactly. If you need a matching set, generate the art only and composite it into one frame you control.
Practical takeaways
- Name the artifact, not the picture: "a collectible card" carries an entire layout with it.
- Let the model pick the illustration style when you have no strong preference; it usually picks a coherent one.
- Spend your effort on the reference photo before the prompt.
- Lock layout first, style second, text last — or skip generated text entirely.
- Keep borrowed visual identities out of anything commercial.
Top comments (0)