--- title: "Storybook City Reel — any story into a cozy hand-drawn street-life reel" tier: hero format: short-form-video theme: illustration | style-capture | storybook | city | social persona: illustration-loving creator, city/tourism account, lifestyle brand, Xiaohongshu/TikTok aesthetic channel duration: "1 story → ~50s vertical reel in ~15–30 min" budget_usd: "$1.50–$5.00 per reel (5 keyframes + 5×10s i2v + 1 music track)" caps: ["gpt-image", "nano-banana", "ltx-q-i2v", "seedance-i2v", "music", "ffmpeg-concat", "ffmpeg-export", "ffmpeg-audio-mix"] skills: ["short-form-video"] showcases: ["/chapters/vancouver-worldcup-storybook-example.html", "/chapters/worldcup-demon-score-example.html"] status: "playbook + flagship showcase rendered E2E (2026-06-10) — Vancouver World Cup opening week; + DEMON-scored variant (soundtrack generated live on music.daydream.live)" reliability: 4.4 # 5 − image gen .3 − i2v .6 − music .2 + .5 proven E2E Vancouver flagship (2026-06-10); 5-stage happy path, assemble + gates deterministic ffmpeg/ffprobe --- # Storybook City Reel — a captured illustration style that turns any story into a cozy animated short This playbook **captures a specific illustration style** — the "cozy storybook street-life" look you see on viral Xiaohongshu / Instagram city reels (a London street, a Paris corner, all flat 2D folk-art charm) — and turns **any user story** into that style: 5 styled keyframes → 5 gently-animated clips → one ~50s vertical reel with an instrumental jazz bed. The style is locked by a verbatim style block pasted into *every* prompt; the protagonist is locked by a name + fixed wardrobe restated in *every* scene. ## The captured style block (use VERBATIM in every keyframe prompt) > Flat 2D hand-drawn storybook illustration, folk-art naive style. Clean uniform dark-brown ink outlines; flat color fills with subtle paper-grain texture; NO gradients, NO 3D shading, NO photorealism; at most minimal soft shadows. Warm muted retro palette: brick red / terracotta, cream / ivory, slate teal-blue, sage and olive green, mustard ochre, dusty rose pink, charcoal; warm off-white sky. Dense European-style streetscape: tall narrow brick townhouses with white-trimmed sash windows, dormers and chimneys; ground-floor shopfronts with dark hand-lettered signboards, window flower boxes, awnings; vintage cast-iron street lamps. Many small charming people at naive proportions (heads ~1/5 of body): vintage 1940s–60s attire — long wool coats, cloche hats, fedoras, scarves, satchels; rosy single-dot cheeks, simple dot eyes; everyone engaged in everyday actions (strolling, chatting at the café counter, cycling, carrying coffee or flowers). Busy but orderly composition with overlapping vignettes; larger figures in the foreground, street receding behind. Blossom trees frame the edges with confetti-dot foliage; petals drift through the air and dot the ground. Mood: cozy, nostalgic, gentle — everyday life is beautiful. The streetscape clause is a *default*, not a cage — swap "Dense European-style streetscape…" details for your city's landmarks (drawn in the same naive style) and the look still holds. The flagship showcase draws Gastown, BC Place and Chinatown this way. ## The animation grammar (use in every motion prompt) Drifting falling petals · people strolling slowly · subtle coat/scarf sway · steam rising from cups · warm shop windows glowing · gentle parallax. **NO camera shake, NO morphing** — the illustration stays flat-2D while elements move. Add exactly **one concrete character action** per scene (sips the coffee, waves the flag, raises the phone, throws arms up, toasts with friends) so the motion gate has something to measure. ## How the agent should run this (interaction contract) 1. **CONFIRM (one message, ≤1 question):** restate in 1 line ("your story as a 5-scene storybook city reel, 9:16 ~50s, instrumental jazz bed, ~$3.20–4.30, ~15–30 min") + real choices: gpt-image (most obedient flat-2D, ~2–4 min/frame) vs nano-banana (much faster, slightly looser palette) · your city's landmarks vs the default European streetscape. Ask only if the protagonist's name/wardrobe isn't derivable from the story. 2. **PREVIEW CHECKPOINT:** the 5-beat script + named cast with fixed wardrobe ($0), then the 5 keyframes (~$0.30–1) gated for flat-2D style + wardrobe hold BEFORE the ~$2.80 i2v pass fires. 3. **NARRATE:** keyframes ≤3 in parallel (~2–4 min each — say gpt-image is slow up front); i2v ≤4 in flight (~2–4 min each); poll `get_create_media` every 10–15s, 1-liner per wave, never silent >2 min. 4. **FAIL GRACEFULLY:** gpt-image client timeout → re-fetch, don't re-render (the image usually landed on fal); ltx-q-i2v 422 → check `resolution:"auto"` + `num_frames:241` (its `duration` param is IGNORED); 3D-ifying/face-morph at the motion gate → re-roll once with "flat 2D stays flat" emphasized → swap to `seedance-i2v` (which needs `resolution:"1080p"`, NOT "auto" — the schemas disagree). ≤2 retries per clip; ship the best passing takes. 5. **DELIVER:** ffprobe-gated reel (audio ≈ video length, 1080×1920) + one honest line ("wardrobe held ~90%; the crossbody bag drifted in scene N — a per-character LoRA is the 95% path") + ONE next step (reuse the style block verbatim on your next story). ## The cap chain ``` any user story → 1. SCRIPT agent breaks it into 5 beats; protagonist gets a NAME + fixed wardrobe → 2. KEYFRAMES gpt-image, image_size:"portrait_16_9" (9:16) # style block VERBATIM + beat + name/wardrobe restated → 3. MOTION ltx-q-i2v per scene — resolution:"auto", num_frames:241, frames_per_second:24, generate_audio:false (= 10s) → 4. MUSIC music cap — lyrics_prompt:"[Instrumental]", warm spring jazz style prompt → 5. ASSEMBLE concat → 1080x1920 30fps yuv420p → loop the music under the full reel → ffprobe gate ``` ### 1 · SCRIPT — name the cast, fix the wardrobe Break the story into **5 beats** (arc: opening moment → rising energy → centerpiece → peak emotion → warm close). Define the recurring protagonist with a **NAME and a fixed wardrobe** — e.g. *"Mei, a young Chinese woman in her early 20s, black bob haircut, mustard-yellow wool coat, white sneakers, small sage-green crossbody bag"* — and **restate the full name + wardrobe in every scene prompt**. Named cast = consistency (named → 0.85–1.0 hold, unnamed → drift; same finding as the novel-recap and emotional-microstory tests). Pick wardrobe colors *from the style palette* (mustard ochre, sage green…) so the protagonist belongs to the illustration. ### 2 · KEYFRAMES — gpt-image with the style block verbatim `gpt-image` (openai/gpt-image-2) has the best stylization fidelity on BYOC — it actually obeys "NO gradients, NO 3D shading" where flux-dev drifts photoreal. **Every scene prompt = the full style block verbatim + the scene beat + the protagonist's name/wardrobe restated**, plus a tail like *"Vertical 9:16 storybook poster composition. Absolutely flat 2D illustration; no photorealism, no 3D rendering."* Pass `image_size: "portrait_16_9"` — **gotcha (verified 2026-06-10):** fal's gpt-image schema takes an *enum* (`portrait_16_9`, `portrait_4_3`, `square_hd`, …), and 422s on pixel-dimension strings like `"1024x1536"`. Expect ~2–4 min per frame; render ≤3 in parallel. **Fallback: `nano-banana`** — also strong on flat illustration, much faster, slightly less obedient on palette. **Gate each keyframe before animating:** open it and check (a) flat-2D storybook per the block — reject any photoreal/3D drift and regenerate with the style constraints emphasized; (b) the protagonist is recognizable with the SAME wardrobe in all 5. ### 3 · MOTION — ltx-q-i2v, with today's cap gotchas `ltx-q-i2v` (fal-ai/ltx-2.3-quality/image-to-video) is the right tier — it animates flat illustrations without 3D-ifying them. **Cap gotchas (verified 2026-06-10, put these in your params):** - `resolution: "auto"` — the fal schema changed; `"1080p"` now 422s on ltx. (This silently broke all ltx i2v until fixed.) - **`duration` is IGNORED by ltx-q-i2v.** To get 10s clips pass `num_frames: 241, frames_per_second: 24` explicitly. - `generate_audio: false` — you're laying a music bed; baked-in audio just fights the mix. - If you swap in `seedance-i2v` as the alternate: it's the *opposite* — seedance **requires** `resolution: "1080p"` (rejects `"auto"`). The two caps' schemas disagree; don't copy params between them. Motion prompt = the animation grammar above + the scene's one concrete character action. **MOTION GATE:** extract frames at t=2s and t=7s — they must show **visibly different poses/action** (petals moved, the arm raised, the sip taken — not just a zoom), and identity/style must hold (still flat-2D; reject any clip that 3D-ifies or morphs the character). ### 4 · MUSIC — instrumental only `music` cap (minimax), **`lyrics_prompt: "[Instrumental]"`** — without it you get random vocals over your cozy street. Style prompt to match the reference reels' "spring jazz" vibe: *"warm spring jazz, brushed drums, upright bass, light piano, cheerful café mood"*. Ask for ~the reel length (≈50s); minimax tracks run ~60s, the loop+trim in step 5 makes the exact length a non-issue. ### 5 · ASSEMBLE — concat, conform, loop the bed, gate Local ffmpeg is fine (or `ffmpeg-concat` + `ffmpeg-export` + `ffmpeg-audio-mix` caps for the server-side path): ```bash # concat the 5 clips (concat demuxer), then conform + mux with looped music ffmpeg -f concat -safe 0 -i list.txt -c copy base.mp4 ffmpeg -i base.mp4 -stream_loop -1 -i music.mp3 \ -vf "scale=1080:1920:force_original_aspect_ratio=increase,crop=1080:1920,fps=30,format=yuv420p" \ -map 0:v -map 1:a -c:v libx264 -crf 18 -c:a aac -b:a 192k -shortest reel.mp4 ``` `-stream_loop -1` on the audio + `-shortest` = the cap-side `audio_fill:"loop"` semantics — the bed runs under the FULL reel, no silent tail. **ffprobe gate before shipping:** audio duration ≈ video duration (±1.5s), ~50s total, 1080×1920, and 3 sampled frames still show the style + the action. ## When to use this playbook - A city / travel / "week in my life" story that wants the **cozy illustrated reel** aesthetic instead of photoreal. - Event tie-ins (a festival, an opening week, a season) where charm > realism and brand-safety matters (everything is drawn, generic flags/jerseys are easy, logos are simply never drawn). - Any brand that wants a *consistent captured style* across many stories — the style block is the asset; reuse it verbatim. ## Honest limits - **Text-in-image is unreliable.** Keep signboards SHORT and generic ('MAPLE HOUR', 'CAFÉ') — one or two words usually render in this hand-lettered style; full sentences won't. Anything load-bearing goes in post. - **Style drift on i2v is the #1 risk.** ltx-q occasionally adds depth/lighting that reads as 3D, or morphs the dot-eyes face. The motion gate exists for this — reject and re-roll with "flat 2D illustration stays flat" emphasized. Ship the best passing take; don't ship a morph. - **Wardrobe ≈ consistency ceiling.** Name + restated wardrobe holds ~85–95% across 5 scenes; tiny accessories (the crossbody bag) drift the most. A per-character LoRA is the 95%+ path if this becomes a series. - **Crowd scenes dilute the protagonist.** Always say "in the foreground" for your hero, or the i2v model animates the crowd and forgets the character action. - **gpt-image is slow** (~2–4 min/frame) and occasionally times out client-side while the image still lands on fal — retry the *fetch*, not the render, before paying twice. ## Cost estimate | Step | Cap | Est. | |---|---|---| | 5 keyframes | gpt-image portrait_16_9 | ~$0.30–$1.00 | | 5 × 10s motion | ltx-q-i2v @ ~$0.056/s | ~$2.80 | | 1 music track | music (minimax) | ~$0.10–$0.50 | | assemble | local ffmpeg | $0 | | **Total** | | **~$3.20–$4.30 / reel** | ## Cross-surface (MCP / CLI / Webapp) - **MCP:** `create_media` ×5 (`model_override:"gpt-image"`, the style-block prompt, `image_size:"portrait_16_9"`) → `create_media action:"animate" model_override:"ltx-q-i2v"` ×5 (`source_url`, `resolution:"auto"`, `num_frames:241`, `frames_per_second:24`, `generate_audio:false`) → `create_media action:"music"` (`lyrics_prompt:"[Instrumental]"`) → `ffmpeg-concat` → `ffmpeg-export` (1080×1920) → `ffmpeg-audio-mix` / `ffmpeg-mux` with `audio_fill:"loop"`. - **CLI:** `livepeer media create --model gpt-image --prompt "