Generative Illustration

Generative Illustration

Make expressive, concept-driven pieces by combining AI-generated illustration assets with code-driven motion. This is the lane designers call 动态平面 / motion graphic design — built the agent-native way: an AI makes the visuals, code gives them life and meaning.

When to use

• An animated illustration, a poetic / ambient "motion poem", a visual metaphor.
• A short brand or art piece where a beautiful illustrated scene moves with intent.
• Any "AI-made visuals + code motion" hybrid.
• Not for: data viz → use chart-animation; pure headline text → use kinetic-typography; 3D / GLSL / heavy particle systems → use webgl-animation.

Setup / prerequisites

• Rendering needs no API key — Node + playwright-core (auto-fetches Chromium) + ffmpeg, all local and free. This covers the compose + animate + export core.
• AI generation (assets + music) needs one API key. The simplest path is a single OpenRouter key — set OPENROUTER_API_KEY — because one key covers all three generators this skill uses:
  • vectors → recraft/recraft-v4.1-vector
  • painterly textures → google/gemini-2.5-flash-image
  • music → google/lyria-3-clip-preview (audio needs stream: true)
• Optional alternatives: REPLICATE_API_KEY for cheaper Recraft (recraft-ai/recraft-v4.1-svg, ~$0.04 vs $0.08/image); a Google AI Studio GEMINI_KEY for Lyria direct.
• Or bring your own assets/music (hand-drawn, any AI tool) and skip generation entirely — the motion + craft core requires no key at all.

The one rule that prevents 90% of failures

Never try to animate one big AI-generated illustration. Generate many small independent elements and let code animate each as a whole unit.

AI image/vector models output a flat pile of ungrouped paths — no <g>, no ids, no "this is the arm / these are the petals." You cannot rig a part inside it. Two failure modes follow:

• Treat the whole image as static and crossfade → looks like a PowerPoint.
• Auto-cluster the paths by proximity and wiggle each → uncanny (motion with no intent, everything wobbles).

The structure lives between assets, not inside one image. Each element is its own clean file; code moves it whole. (This is how working motion-illustrators do it — design the pieces separately, then code organizes and presents them.)

Pipeline

1. Concept — one metaphor, ruthless restraint.
2. Generate each element separately with AI, isolated on a plain background.
3. Strip the background → transparent sprite.
4. Code the motion — a deterministic seek-harness on a <canvas>; place + animate sprites + text + grain.
5. Render — frame-grab every frame → ffmpeg → MP4.
6. Music — generate a custom track (Lyria), mix it in.

1 · Generating assets

• Crisp vector elements (a flower, an icon, a creature) → Recraft v4.1 SVG.
• Painterly textures (a watercolor wash, paper grain) → an image model (Gemini "nano-banana").
• Specificity wins. Name a real, characterful subject — generic prompts give generic art:

• Always add no text — AI mangles lettering. Add every word in code (where it is exact and editable).
• Ask for "single isolated element, plain solid white background" so it is trivial to cut out.
• Strip the background path. Recraft draws a full-canvas white rectangle as path #0. Detect the path whose bounding box covers ~the whole viewBox and delete it → transparent sprite:

# keep every path EXCEPT the one whose bbox spans the whole canvas
def is_bg(path, VBW):
    nums = [float(n) for n in re.findall(r'-?\d+\.?\d*', d_of(path))]
    xs, ys = nums[0::2], nums[1::2]
    return xs and (max(xs)-min(xs)) > VBW*0.92 and (max(ys)-min(ys)) > VBW*0.92

• Recraft v4.1 SVG format: <path …></path> (not self-closing), viewBox="0 0 2048 2048", transform="translate(x,y)", absolute coordinates.

2 · The deterministic seek harness (the key to clean rendering)

Every visible value is a pure function of time t — never a requestAnimationFrame loop or a CSS transition (those desync from the frame-grabber and flicker).

window.__T = totalSeconds;                 // total duration
window.__seek = (t) => render(t);          // bar width, x/y, opacity, scale — all derived from t
// on load: render the frame named in the URL
window.__seek(parseFloat(new URLSearchParams(location.search).get('t')) || 0);

A one-browser Playwright loop seeks t = 0, 1/30, 2/30 …, screenshots each, then ffmpeg assembles them (use even dimensions + yuv420p). For stateful sims (particles), step the sim forward deterministically with a seeded RNG so every render is identical.

3 · Animating sprites

• Load each asset as an Image (inline base64 data-URI so the HTML is self-contained); drawImage per instance inside save/translate/rotate/scale/restore.
• Anchor each sprite at its motion pivot (base / center / a corner — whatever the motion needs). E.g. to rise from a line: translate(x, baseY) then draw at (-w/2, -h).
• Match the entrance to the idea, don't reflex to "grow": a count-up, a split/double, a slide-in, a flood, a snap, a scale-from-0 — each says something different. Idle life: gentle sin(t) drift.
• Add a grain overlay (seeded noise) for a hand-made, Softcore texture.
• In capture, gate on an window.__ready() flag once all images have decoded.

4 · Craft — what makes it good, not just moving

• Let the concept choose the visual — never default to one device. The motion vocabulary must be derived from this idea, not reused: compounding → a dot doubling until it floods the screen (not a plant); a blank page → the first word breaking a dam of text (not a sprout); overload → tabs/windows multiplying into noise. Reusing one device — "a thing grows up from the ground" — across unrelated concepts is the #1 reason these go stale and generic. If you reach for "something grows" a second time, stop and ask what the concept actually looks like. (The growing-plant in the original garden piece worked only because the concept was literally botanical.)
• Causality. Tie cause → effect literally (an input lands → something happens at that spot). Loosely-related parallel events read as random noise.
• Restraint. Motion needs fewer elements than static design — too many and it gets noisy in motion. Keep only the core; don't overdo it.
• Spotlight the hero. When one element matters (the one that "made it"), dim the rest and lift it — a halo, a scale, a color shift.
• Rhythm: resolve → a beat of stillness → reflect. A closing line is a looking-back statement; it only lands after the action has stopped. Let the motion finish, hold a beat, then reveal the slogan — never overlap the conclusion with things still moving.
• Easing. Never linear on anything a viewer reads. Ease the value, not just opacity. Overshoot for small pops; slow ease-out for growth.
• Semantics — don't self-sabotage. If the piece depicts failure/withering, do not label it with your own product's features (it reads as "our product fails at its core jobs"). Use neutral or process language instead.

5 · Music

Generate a custom track with Lyria 3 (OpenRouter google/lyria-3-clip-preview). It is AI-generated, so it is royalty-free. Audio output requires stream: true — chunks arrive as base64 in delta.audio.data; concatenate and decode to MP3. Mix with ffmpeg: trim to length, fade in/out, duck the volume (~0.5) under any narration.

6 · Verify loop (do not skip)

Render the piece, build a full-duration contact sheet (sample frames across the whole timeline, not just the first seconds), and read it yourself: blank frames? overlapping rows? does the concept actually read? Fix and re-render. This deliver-and-verify loop is what separates "it ran" from "it's good."

references/

• asset-pipeline.md — asset prompting, background-strip code, Recraft/Gemini format notes
• render-harness.md — full seek-harness HTML + Playwright capture + ffmpeg recipe
• craft.md — causality, rhythm, restraint, spotlight, with worked examples

The craft behind iart.ai. Cost reference: a Recraft SVG ≈ $0.04, a Lyria track ≈ cents — a whole finished piece comes in under $1.