Лис Микита · FIFA Final Cup

The Final Set (v4)

Seven match-day logos generated by feeding the bar's original logo files to Google's Gemini image model as style references — so the output is the same artist, not an imitation. Scroll on for how it got here.

The Source: Three Reference Logos

Everything starts from these three originals — the bar's actual logo set. The system they encode: one fox (long sly snout, feathered kozak hat, embroidered folk vest, black britches, cream leg wraps), one flat burnt-orange background, one lockup (black Cyrillic banner + thin serif), and exactly one themed swap per variant — a Santa hat, a soccer ball, a cane.

The Timeline — Every Generation, Failures Included

Four batches in one afternoon (July 18, 2026). The first three tried to describe the style in words to xAI's Grok image model; each prompt "envelope" got closer. The fourth stopped describing and handed the model the originals.

Gen 1 · "FIFI" Two strikes

Grok (grok-imagine-image) · text-only style envelope, first draft · ~$0.10/image

• Strike one — the typo. The request said "FIFI" and the robot dutifully printed FIFI FINAL CUP on all seven images without asking. (Lesson logged: flag suspicious proper nouns before spending tokens.)
• Strike two — the style morph. The envelope described "Ukrainian folk woodcut" in words, and Grok answered with a heavier, busier illustration: dense crosshatched fur, full sleeved team jerseys, a chunkier fox. Handsome — but a different artist.

Also from this batch: flag cape · Argentina kick · trophy

Gen 2 · The Overcorrection Right words, wrong weight

Grok · envelope v2: "lanky fox, bare arms, minimal hatching, four flat inks" · ~$0.10/image

• FIFA spelled right this time, team colors moved to scarves and sashes instead of jerseys — both correct calls.
• But the envelope said "sparse embroidery, minimal hatching" and Grok swung all the way to a thin, uniform, cute pen line. The references are a woodcut badge — heavy black mass, carved strokes. v2 reads as an indie children's book.
• Also lost: the feathered kozak hat (it's canonical — the reference fox wears it even mid-kick).

Rest of the batch: matador · flag cape · Argentina kick · trophy

Gen 3 · The Woodcut Envelope Best text-only attempt

Grok · envelope v3, written from a side-by-side critique of v2 vs the references · ~$0.10/image

• The critique named six specific misses and the envelope pinned each one: heavy black ink mass, chunky carved linework with white knife-marks, the feathered hat restored, dense embroidery, toothy sly grin, fox planted on the banner.
• Result: genuinely close. The mate-break panel is the best text-prompted image of the day.
• Remaining drift you can't fully prompt away: some outputs came back landscape instead of portrait, "kicking" kept rendering as dribbling, and the trophy came out bowl-shaped. This is the ceiling of describing a style in words.

Rest of the batch: Spain kick · matador · Argentina kick

Gen 4 · Stop Describing, Start Showing The unlock

Reference-image APIs: OpenAI gpt-image-1 (images/edits) vs Google gemini-2.5-flash-image · originals attached

• Both APIs accept the actual reference PNGs alongside the prompt. A head-to-head on the same subject (flag cape) settled it in one round.
• OpenAI: kept the vibe, redesigned the character — rounder sports-mascot face, different banner, different artist. Gemini: indistinguishable from the original logo family. Winner.
• With references attached, no style envelope is needed at all — prompts shrink to one-sentence scene descriptions ("change only the pose, the props, and the small text line").

Gen 4½ · The Spelling Saga ARGANTINA · ARGNTINA · ESPÃNA

Gemini re-rolls, Gemini edit passes, and finally a deterministic local patch

• Gemini could draw the artist perfectly but could not spell ARGENTINA — three images shipped with three different misspellings, and edit passes kept anchoring on the wrong text already in the image.
• The fix ladder: (1) re-roll with "must be spelled exactly" — fixed one, and upgraded the matador to a proper accented ESPAÑA; (2) a Gemini edit pass naming the misspelled word — fixed two more; (3) for the unkillable versus card, a Python script: locate the text band by row-darkness profile, paint it out with the sampled background orange, re-typeset the line in Didot (the closest system font to the lockup serif), width-fitted and letter-tracked.
• Every misspelled intermediate was kept. Receipts below.

Full evidence chain: versus №1 · №2 · №3 · №4 · trophy №1

Ship It v4 final

Seven clean logos, correctly spelled, same artist as the originals

The full set is at the top of the page. Total spend across all four generations, both head-to-head tests, every re-roll and every repair: roughly three dollars.

How It's Made — The Pipeline

Everything is reproducible from three small shell scripts and two TSV files that live next to the images. Delete any image, re-run the batch, and only that one regenerates.

What the afternoon taught

• Text envelopes approximate; reference images match. Three prompt drafts got asymptotically close to the house style. One API call with the originals attached was the house style.
• Pick the model per job, not per loyalty. Grok is a fine improviser, OpenAI a fine mascot designer, Gemini the only one that impersonated the original artist — and none of them can reliably spell ARGENTINA.
• Keep the failures. Every bad batch, misspelling, and dead-end test file stays in the tree with notes. That's what makes this page — and the next variant — possible.
• Typos are load-bearing. Gen 1 shipped as "FIFI FINAL CUP" because a human typo'd and a robot obeyed. Now it's policy: suspicious proper nouns get flagged before generating.