Our running lab journal for the chiva book's art direction. Every experiment we run gets an entry here — model, what we tried, verdict, score — so we never re-test a dead end.
Tried: Edit pass to add driver/munchos/pig, then scripts/overlay.ts stamped "2+1=3" onto the blank banner (hand-lettered font, slight rotation).
Verdict: Overlay text looks genuinely hand-painted — the text pipeline is proven. But the edit pass recolored the chiva (red→teal) and drew pig-snouted goblins ("pig" bled into the creatures). Need characters right in the T2I pass instead.
Tried: FLUX.2 with blank banners — all text to be overlaid programmatically afterwards.
Verdict: The most Nordqvist-like art so far: loose ink, bleeding washes, calm sky, coffee terraces. But it dropped ALL characters (no driver, no munchos, no pig).
Tried: Model shootout winner: true watercolor, paper grain, muted palette. Cleanup edit pass removed its hallucinated pseudo-text, stray signs and license plate.
Verdict: Best chiva ("LA AVENTURERA" lettering, folk paint) and best composition. Style slightly busier/more saturated than Nordqvist. Menos & Más ended up walking instead of aboard. Strong candidate.
Tried: Model shootout with the loose-watercolor style prompt.
Verdict: Charming flat-gouache look, best text rendering of all, but ignored 16:9 (returned 1:1) and not watercolor. Good backup if text-in-image is ever wanted again.
Tried: Model shootout with the loose-watercolor style prompt.
Verdict: Still digital cartoon; hallucinated a second vehicle and a wrong sum. Out.
Tried: Fed frame 1 back as its own reference to refine it.
Verdict: Beautiful chiva paintwork and distinct muncho characters — but lost the colonial plaza, the church, and the marranito. Kontext edits rather than follows the full brief.
Tried: Scene first, explicit "no train no rails", chiva description strengthened.
Verdict: Proper chiva + colonial plaza + correct "2+1=3", but ghost rails under the wheels and munchos read as human kids. Still glossy.
Tried: First attempt, style block referencing "Muckletåget".
Verdict: Drew a literal train on rails with generic cute animals. The word "Muckletåget" (the Muckle TRAIN) primed trains. Style glossy digital. Lesson: never name the original book in the prompt; scene must come first.
Tried: ByteDance Seedream 5.0 Pro (endpoint: bytedance/seedream/v5/pro — no fal-ai prefix), blank banners, programmatic "2+1=3" overlay afterwards.
Verdict: Best frame yet: loose genuine watercolor, colonial street, chiva with folk paint, driver with mustache and sombrero vueltiao, two striped munchos actually ABOARD, marranito running with its neckerchief, and a blank banner placed perfectly in the tree for the overlay. No stray text at all — the --notext rule was obeyed completely. Current champion.
Tried: Same frame-1 prompt but text ALLOWED (no --notext), to test whether Seedream 5 can paint the sum itself. New naive-character steering active.
Verdict: Best image so far. "2+1=3" legible and hand-painted on the banner, "LA AVENTURERA" correct on the chiva, munchos noticeably simpler/more stylized (bean bodies, hats) thanks to the steering. Decision: let the model paint the sums; scripts/overlay.ts stays as fallback; narration text is always added deterministically with scripts/pagetext.ts. Caveat: one sample — watch text reliability on harder sums (e.g. "18-2=16").