Art-directed graphic

Why this exists

Nearly every AI holiday/greeting graphic dies the same death, and it is diagnosable, not a matter of taste. The failure signature:

• soft pastel gradient wash background
• a symmetrical, centred photoreal still life
• glossy specular highlights (honey, glass, fruit)
• everything smooth, evenly lit, no surface, no edges, no ink
• typography parked ON TOP of a finished render rather than designed into it

That combination is what everyone else's post looks like this week. The client sees "AI" before they see the idea.

The fix is not a better render. It is a different medium. Move to something with a physical surface - torn paper fibre, screen-print halftone, ink misregistration, canvas weave, carved linocut marks, thread, glaze crazing. Diffusion models render those extremely well, and they are exactly what slop lacks. Surface is the tell.

The method, in four moves

1. Name a medium, not a subject

A prompt that describes what is in the picture gets slop. A prompt that describes what the picture is physically made of gets a design. Every prompt must name:

• the medium ("a four-colour silkscreen on aged uncoated paper", "a handmade cut-paper collage photographed flat under raking north daylight")
• its imperfections - this is the load-bearing part. Ink misregistration with a 1mm offset. Torn paper fibre at every edge. Dry-brush drag and gouache pooling. Visible canvas weave. Pencil under-drawing showing through. Hand-cutting asymmetry.
• a strict palette, counted ("exactly four inks: crimson, warm gold, deep ink navy, and the cream of the paper"). Say the number. Unbounded palettes go muddy.
• a lighting/finish rule ("completely flat lighting, no gradients, no photographic realism" or "soft raking light so the cut paper lifts and casts delicate shadows")

See directions.md for the library of media that work, what to name in each, and where each one fails.

2. Generate the art CLEAN - no text, no logo

Both are composited afterwards, always:

• Non-Latin scripts garble in-model (Hebrew, Arabic and others). Even when a model gets Latin right, it will not get the client's font, tracking or line breaks.
• A generated logo reads pasted-on every single time. Key the real file.

Instead, reserve space in the prompt, positively phrased:

"The upper third of the image is left deliberately quiet and empty as reserved space for typography that will be added later. The image is purely pictorial and contains no writing, letters or numerals of any kind."

If the model puts a dark plane in the reserved zone anyway, do not fix it with a scrim - the scrim turns it into a grey smear. Add an explicit composition rule and regenerate:

"COMPOSITION RULE: every painted plane, collaged paper and contour line sits entirely within the lower two thirds. The whole upper third is one single unbroken field of plain warm cream painted ground, running uninterrupted from the left edge to the right edge and up to the top edge."

3. Composite typography locally, in HTML + headless Chrome

Not Pillow. Chrome gives real RTL, real font shaping, real flexbox layout, and the whole lockup becomes free to iterate - the expensive generation happens once, copy and layout revisions cost nothing. This matters because clients revise copy 4-5 times.

4. One artwork, many formats

Design a print master and crop the social sizes from it, or generate one square "plate" and let both formats draw on it. Never design the same graphic twice.

Layout system that works

Print geometry: A5 is 148×210mm. With 3mm bleed all round that is 154×216mm, which is 1819×2551px at 300dpi. Nothing important within ~8mm of the trim edge. Ship a bleed PDF and a trimmed PDF; ask the printer which they want and whether they want crop marks (some trade printers prefer no marks).

Closest generated ratio to A5 is 2:3 - crop the rest with object-fit: cover.

Hard-won gotchas

• Check your image provider's exact endpoint and parameters. Some models take aspect_ratio plus a resolution tier and silently ignore image_size. High-resolution generations can take a minute or more, so make generation resumable.
• Board-style prompts come back with a white margin around the "photographed" artwork. Auto-crop to the painted ground (detect the warm-vs-neutral edge) before use.
• Headless Chrome is OOM-killed on 1819×2551 pages when renders overlap, and silently leaves a stale PNG. Retry each screenshot up to 3×.
• Logo keying: key the white paper out by luminance, keep the real colours for light grounds; make a flat gold/white monochrome version for dark grounds.
• Negatives summon the subject. Never write "no candlesticks" - describe the scene positively instead. The one safe negative is "no writing of any kind", which Gemini-class models honour.
• No em-dashes anywhere, in any language. No exclamation marks on greeting cards (a style choice, adjust to the client).
• Palette discipline beats everything. If a direction comes back muddy or collapsed into one drab tone, it is a regenerate, not a colour-correct.

Getting the brand INTO the art, not onto it

The difference between a design and a stock image with a logo: the brand idea should be inside the picture, so the graphic is not logo-swappable.

Before prompting, extract from the client:

1. What the logo mark actually depicts (for example a globe with meridian arcs: a globe or meridian motif can recur inside the artwork)
2. The client's world-objects (cargo ship, aeroplane, crates, route arcs)
3. A directional idea (for example routes that always CONVERGE on one place and never radiate out, because the copy angle is "coming home")
4. The brand colour, as an accent only - the occasion's palette leads

Best-in-class device: hide the brand objects inside ornament so finding them is a reward for looking (a ship and aeroplane concealed in papercut filigree).

Suggested job layout

Keep one folder per job with four small pieces so a revision costs nothing:

• prompts.py: the prompts, with the reasoning in the docstring
• gen.py: image-generation calls, resumable (skips art that already exists)
• build.py: copy, palette and layout per direction, writes HTML
• render.sh: headless Chrome screenshots with retry

Per-client taste

If the client has a taste-calibration note, read it before prompting. After every round append what was liked, what was struck out, and the reason.