Anime Animation Production

Anime and stylized 2D animation production

Anime is not a preset. It is a set of production conventions born from hand-drawn cel animation under tight schedules: flat color, hard-edged shading, deliberately uneven frame timing, and a shot grammar that spends motion where it matters and freezes everywhere else. Generative tools can reproduce the surface of that look easily, and the grammar almost never by default. The value an agent adds is the grammar — knowing which convention a given genre and beat calls for, encoding it into prompts and shot plans, and reviewing output against the things that actually break the illusion.

This skill covers the craft as production knowledge. It is provider-independent. Any model named (Sora, Kling, Vidu, Wan, PixVerse, Midjourney, etc.) is an illustrative option, dated where its behavior is volatile — never the method itself.

Three labels used throughout

To keep evidence honest, claims here are tagged:

• [Fact] — documented in provider docs, animation references, standards, or news of record, with a source and (for volatile items) a verification date.
• [Convention] — a craft practice established in professional 2D/anime animation, independent of any tool.
• [Heuristic] — a production rule of thumb for generative tools, empirical and model-dependent; verify against your actual model and re-test when models change.

Volatile facts in this document were verified on 2026-07-10. Model capabilities and platform policies change fast; re-verify before relying on any dated claim.

When to use / when not to use

Use this skill when the deliverable should read as anime or stylized 2D animation — cel-shaded characters, limited-animation feel, anime shot grammar — regardless of the underlying model.

Do not reach for this skill for photoreal or 3D-CG-realistic pieces, for Western cartoon styles (Disney full animation, Cartoon Network flat-vector, cutout/Flash), or for generic "illustrated" video that has no anime intent. Those are different craft traditions with different timing and shading logic; applying anime conventions to them produces an uncanny hybrid. If the user says "animated" without specifying anime, confirm the tradition before committing to this grammar.

Part 1 — The visual grammar of anime as production knowledge

1.1 The cel look: line and flat color

The defining surface of anime comes from the celluloid process: an ink outline drawn on a clear cel, backed by flat painted color, over a separately painted background. [Fact — the "cel" in cel-shading names the celluloid sheet used in traditional production; painters laid flat color behind ink outlines. Sources: Wave Motion Cannon; cel-shading references, 2026-07-10.] Its production-relevant properties:

• Clean, weighted linework. A continuous outline of controlled thickness, usually darker than a pure black (often a brown-black or color-holds on hair/skin). Line weight varies by era: heavier, chunkier lines in Akira-era/OVA work; thinner, more uniform lines in modern digital TV anime. [Convention]
• Flat color fills with hard shadow boundaries. No soft airbrush gradients on the character. Shading is one or two discrete tones with a crisp edge, not a blur. A second shadow tone ("2-shadow") appears in higher-budget or dramatic work. This hard-edge cel shadow is the single strongest tell of "anime" versus generic digital painting. [Convention]
• Specular hair highlights. Chunky, geometric highlight blocks on hair (not soft sheen). [Convention]
• Painterly backgrounds against flat characters. The character/background contrast is deliberate: characters are flat cel, backgrounds are lush painted (watercolor, gouache, or digital matte in the Kanmanagement/Kobayashi tradition). This flat-on-painted contrast reads as "anime" even in a still. [Convention]

Production implication: if your output shows soft, blurred, airbrushed shading on faces, or lineless "soft render," it has drifted toward generic digital illustration and away from cel. That is a QC failure (see Part 8), not a style choice, unless the brief explicitly wants a painterly/"Makoto Shinkai" soft look.

1.2 Limited animation and framerate modulation

Anime evolved limited animation — fewer unique drawings per second than full Western animation — as an economic necessity that became an expressive language. This is the most important and most-ignored convention in AI anime.

Frame timing vocabulary [Fact — Anime News Network lexicon; Wave Motion Cannon, 2026-07-10]:

• [Fact/Convention] TV anime is broadly animated "on threes" (~8 unique drawings/sec) as a production standard adopted after the 1960s; Osamu Tezuka's Astro Boy (1963) pioneered limited techniques under budget pressure.
• Framerate modulation = changing the rate within a single cut: a thrown weapon snaps on ones, a lumbering giant crawls on threes or fours, an impact holds a single drawing for several frames. Different rates carry different weight and speed; the brain fills the gaps. [Convention — popularized through the Otsuka/Kanada lineages.]
• Element-wise rates. In one shot a fast foreground effect can run on ones while a large slow object runs on twos, steering the eye. [Convention]

Production implication for generative video: most AI video models interpolate to smooth, high-frame-rate, "on-ones-everywhere" motion — which reads as too fluid, uncanny, "AI-smooth," and un-anime. Countermeasures [Heuristic, 2026-07-10]:

• Ask the model for "limited animation," "animate on twos/threes," "stepped animation," "held frames," "8fps anime cadence," "TV anime motion." Effectiveness varies by model; many still smooth it out.
• More reliable: post-process to reduce effective frame rate. Render at the model's native rate, then step/decimate frames (hold every drawing 2–3 output frames) or render fewer keyframes and hold. This gives the stepped cadence that prompts alone usually cannot. [Heuristic]
• Reserve genuine on-ones fluidity for one or two sakuga moments; make everything else more limited so the fluid beat lands.

1.3 Sakuga, impact frames, smears, and the effects vocabulary

• Sakuga (作画, "drawing pictures"): in fan usage, a standout burst of highly fluid, detailed, virtuoso animation — the money shot. [Fact — term/definition per animation references, 2026-07-10.] Producing it means spending your fluidity budget on a short beat while keeping surrounding shots limited, so it contrasts.
• Impact frame: at the moment of a hit, one or two extreme frames — often a high-contrast, inverted, monochrome, or abstract "flash" drawing — inserted for 1–3 frames to punch the collision. [Convention] In generation, produce as an inserted still or a hard cut, not as smooth motion.
• Smear frames: deliberately distorted, elongated in-between drawings that only exist for a frame or two to sell speed (a hand becomes a blur of multiple fingers). [Convention]
• Speed lines / motion lines: graphic straight or radial lines overlaid to indicate velocity or focus (radial "focus lines" for emphasis/reaction). Structural to the medium, not decoration. [Convention]
• Effects animation (FX): fire, water, smoke, energy, and debris in the stylized "Kanada style" (angular, geometric flame/lightning shapes; "wakame"/seaweed-shaped shadows). [Convention — Kanada lineage.]
• Bank shots / held cels: reused animation (transformation sequences, stock attacks) and long holds on a single drawing with only a camera move or a mouth flap. [Convention] The mouth-flap-over-a-hold and the slow pan/zoom over a still are the workhorses of TV anime economy and are easy to reproduce with a still + a camera move.

1.4 Screen tone and halftone

From manga heritage: screen tone (halftone dot patterns and hatching for gray values and texture) and halftone impact bursts. In color anime these appear as retro/manga-panel stylings, flashback textures, and comedic cutaways. [Convention] Useful as a deliberate stylistic tag ("halftone screentone shading, manga panel"), not a default.

Part 2 — Genre and demographic styling

Anime "style" is really several styles keyed to demographic and genre. Getting this right is the difference between generic-anime and on-target. Characteristics below are [Convention] for the traditions and [Heuristic] for the prompt tags.

• Shonen (action, young-male-skewed): bold, dynamic poses; heavier confident linework; saturated high-contrast palette; kinetic FX (speed lines, energy, impact frames); exaggerated musculature and expressions. Tags: cel-shaded, dynamic action pose, bold outlines, saturated colors, speed lines, impact frame.
• Shojo (romance/drama, young-female-skewed): large expressive eyes with detailed catchlights; delicate, flowing linework; soft pastel palette; emotive environmental effects (floating petals, sparkles, screen-tone bloom, backgrounds that mirror mood). Tags: soft pastel palette, delicate linework, large sparkling eyes, floral bokeh.
• Seinen (mature, adult-male-skewed): more realistic proportions and anatomy; finer detailed linework and denser shading; muted/desaturated palette with selective accent color; psychologically complex, restrained expressions. Bridges anime and Western illustration. Tags: realistic proportions, muted desaturated palette, detailed shading, subdued expression.
• Josei (mature, adult-female-skewed): grounded, elegant character design; softer realism than seinen; naturalistic palette. Tags: naturalistic palette, refined realistic features.
• Slice-of-life / iyashikei: clean approachable character designs; detailed, realistic everyday backgrounds (Japanese suburbia, classrooms); muted naturalistic light; warmth over spectacle. Tags: soft natural lighting, detailed realistic background, gentle muted palette, everyday setting.
• Mecha: hard-surface mechanical detail, line-heavy panel lines, metallic cel highlights, often combined with dramatic FX. Tags: mechanical detail, hard-surface shading, panel lines.
• Chibi / super-deformed (SD): 2–3 head-height proportions, simplified features, for comedy inserts. Tags: chibi, super deformed, simplified.

Retro-90s / OVA cel look

A distinct and popular target. The look comes from a real production chain: hand-painted cels in a small fixed palette, shot on 16/35mm film, broadcast through CRT. [Fact — 90s production-stack references, 2026-07-10.] Reproduce it with an explicit stack of tags rather than a generic "retro" word [Heuristic, 2026-07-10]:

• Rendering: cel paint, hard-edge cel shadow, chunky hair highlights, heavy line weight (Akira-era), hand-painted background.
• Film/broadcast artifacts: film grain (this reads more "90s" than "VHS," which skews 80s), slight gate weave, warm white point, faded/limited palette (muted red, smoky teal, faded mustard).
• Optional artifacts: chromatic aberration, scanlines — but sparingly; overuse creates the "meme VHS filter" look rather than authentic OVA.

Reference touchstones for the era's feel (as descriptors, not to imitate a specific protected work): Akira, Cowboy Bebop, Evangelion, Sailor Moon-era palettes.

Part 3 — Prompting vocabulary that controls anime style

[Heuristic, model-dependent, 2026-07-10.] Vague "anime style" produces a generic soft-modern-digital look. Control comes from stacking four categories:

1. Medium/technique tag — the single most load-bearing word. Prefer "cel-shaded" (or "cel shading," "anime cel") over "anime style." Add "flat colors, hard-edge shadows, clean lineart."
2. Genre/demographic tag — from Part 2 (shonen/shojo/seinen/slice-of-life/retro-OVA).
3. Palette and lighting — explicit colors and light quality ("muted desaturated palette," "warm sunset rim light," "high-key soft daylight").
4. Line and shading specifics — "bold line weight," "thin uniform lineart," "two-tone cel shadow," "chunky specular hair highlights."

Additional levers:

• Negative direction (where supported): exclude "photorealistic, 3d render, blurry gradients, soft airbrush, lineless" to hold the cel look.
• Camera/lens language still applies ("wide establishing shot," "low-angle hero shot," "shallow depth of field bokeh background") and composes with style tags.
• Do not over-stack contradictory tags (retro OVA grain + hyper-clean modern digital) — pick one coherent era.
• Provider-specific note: some models expose dedicated anime/illustration modes or anime-tuned checkpoints/LoRAs that respond better to booru-style tag lists (comma- separated attributes) than to prose; others (large general video/image models) respond better to natural-language scene descriptions. Match prompt shape to the model. [Heuristic, 2026-07-10.]

Part 4 — Style consistency across shots and episodes

This is the hardest technical problem and where most AI anime fails. Two distinct consistency targets:

• Character consistency (on-model): same face, hair, proportions, costume, and color scheme across every shot. [Fact — character/"identity" drift is documented across essentially all current major video models; subtle changes to face, hair, and proportions appear between shots and worsen over longer clips and scene changes. Sources: multiple 2025–2026 model comparisons, 2026-07-10.]
• Style consistency (whole-look): same line weight, shading logic, palette, grain, and rendering across shots and episodes, so nothing drifts between "flat cel" and "soft painterly" or between two different color temperaturas.

Note [Heuristic]: stylized anime characters tend to drift less than photoreal humans, because exaggerated proportions and high-contrast features are easier for models to reproduce — a genuine advantage of the medium. It is a smaller problem, not a solved one.

Workflow that holds consistency

1. Build a character model sheet first. Generate (or assemble) a reference sheet: front / three-quarter / profile / back, a neutral and one expressive face, and a full-body turnaround, all in the final style. This is the anime-production "settei"/model-sheet concept adapted to generative work. [Convention + Heuristic]
2. Curate a small, diverse anchor set. For per-shot reference-conditioning, use a few distinct angles/expressions rather than many similar ones. [Heuristic — 2–4 well-chosen references typically outperform 5–6 redundant ones; overlap gives the model no new information; 2026-07-10.]
3. Anchor every generation to references. Use whatever conditioning the model offers: image reference / character reference (some models accept multiple simultaneous references), first-frame image, or a trained character LoRA/embedding for repeated use. [Fact — as of 2026-07-10, reference-image and first/last-frame conditioning are available in several video models, e.g. Kling's reference and start/end-frame modes, Sora's image reference / storyboard keyframes, Vidu's multi-reference mode; treat exact feature names/limits as volatile and verify.]
4. Lock a style anchor. Keep one "hero" frame or style reference and condition toward it so the whole sequence doesn't drift between photoreal and illustration or between lighting regimes. [Heuristic]
5. Chain shots. Use the last good frame of shot N as the first-frame reference for shot N+1 where continuity is tight; for episodic consistency, keep the same model sheet, seed strategy, and LoRA across the whole batch.
6. Validate every shot against the sheet before accepting (Part 8). Reject and regenerate off-model shots rather than "fixing in the edit."

Style drift traps [Heuristic, 2026-07-10]

• Adding scene context ("in the rain," "at night," "in a crowd") can silently soften edges, mute colors, or re-render fabric/hair — changing the character's signature. Re-assert the core style/character tags in every prompt; don't assume they carry over.
• Palette drift across a batch is common; specify exact hex-or-named colors for hair, eyes, and signature costume in every prompt, and QC color-pick key frames.
• Longer single clips drift more than short ones; prefer shorter shots stitched in the edit over one long generation when identity matters.

Part 5 — Shot planning and motion direction adapted to anime

Anime editing rhythm is distinctive and generative shots should be planned to it, not to a generic "smooth continuous" video model default.

Cutting rhythm and shot types

• The hold + camera-move shot. A large fraction of TV anime is a still (or near- still) drawing with a slow pan, zoom, or rack, plus a mouth flap or hair sway. This is economical and the model reproduces it reliably: generate a strong still, then apply a slow Ken-Burns/parallax move. Use it for dialogue and establishing beats. [Convention
  • Heuristic]
• Reaction inserts. Anime cuts frequently to a brief static reaction (eyes widen, sweat drop, focus lines) — often 0.5–1s. Build these as short beats; they carry emotion and cover the fact that most shots are held. [Convention]
• The 3-hit action beat. Action often reads as: anticipation pose (hold) → fast in-between/smear (on ones) → impact frame + hold on the connect. Plan action as this spend-fluidity-then-hold cadence rather than uniform continuous motion. [Convention]
• Establishing pillow shots (pan over background). Slow drift across a painted background to set place/mood between scenes (the "pillow shot" / kire). [Convention]

Motion direction levers for stylized action

• Request stepped/limited motion for most shots and fluid on-ones only for the sakuga beat, to create contrast (Part 1.2).
• Direct weight and timing explicitly: "heavy anticipation, then a single fast frame, then hold on impact" reads more anime than "character punches smoothly."
• Add speed lines and impact flashes as overlays or inserted frames rather than expecting the base render to include them convincingly.
• Keep the camera anime-grammar: whip pans, snap zooms to a face, dramatic low/Dutch angles for tension, and long slow pushes for emotion.

Part 6 — Audio pairing conventions

Anime sound is stylized and expressive, and pairing it correctly does as much for the "anime" read as the visuals. [Convention, with sourced anime-sound references, 2026-07-10.]

• Voice (seiyuu tradition). Performances are emotionally heightened and clearly delivered; dialogue sits forward in the mix and stays intelligible over music and FX. For AI/synth VO, direct expressive, character-distinct performances; keep dialogue the priority channel.
• Exaggerated, layered SFX. Impacts are larger than life — a punch layered with a synth kick/tom, an energy blast built from synth + whoosh. Footsteps, cloth, and small motions are often over-emphasized. Layer designed hits rather than using dry realistic thumps.
• Impact stingers on the beat. Sync the loudest hit to the impact frame; a musical or percussive stinger on the connect sells action.
• Ma (間) — expressive silence. Anime deliberately drops music and SFX to near-silence for emotional or tense beats (breath, room tone only). Don't wall-to-wall the music; plan silence as a tool.
• Music as mood/pacing bed. BGM supports rather than competes; it swells for climaxes and drops out for ma. Genre cues matter (city-pop/jazz for a Bebop feel, orchestral for epic shonen, gentle piano/acoustic for slice-of-life).

Mixing order of priority: dialogue intelligible first, then impact hits, then music bed, with silence used intentionally.

Part 7 — Cultural, legal, and platform sensitivities

This is a live, contested area. Handle it explicitly; do not treat AI anime as consequence-free.

Studio- and artist-style imitation

• [Fact, 2026-07-10] Style is not, in itself, protected by U.S. copyright — generating images "in the aesthetic of" a studio is not automatically infringement. But this is a legal characterization of style, not permission to copy characters, and it is disputed across jurisdictions.
• [Fact, 2026-07-10] The "Ghiblification" episode (March 2025): OpenAI's image tools produced viral Studio-Ghibli-style images. OpenAI stated it refuses the style of individual living artists but permits broader studio styles (a policy critics note is in tension, since Ghibli's style is inseparable from living artist Hayao Miyazaki, who has publicly expressed strong opposition to AI animation). Sources: TechCrunch, Fast Company, 2025.
• [Fact, 2026-07-10] Japanese rightsholders are actively opposing training on their work. In late October 2025, CODA (Content Overseas Distribution Association), representing ~18+ major Japanese companies including Studio Ghibli, Bandai Namco, Square Enix, Kodansha, Kadokawa, and Shogakukan, publicly demanded OpenAI stop using their content to train Sora 2, arguing Japanese copyright law implies an opt-in standard and that reproduction during training can itself infringe. The Japanese government/Cabinet Office also formally raised concerns. Sources: Variety, TechCrunch, Game Developer, 2025.

Practical guidance for an agent:

• Prefer original character and world designs and a generic genre style ("cel-shaded shonen action," "90s-OVA look") over "in the style of [named living studio/artist]." Naming a living studio/artist as a style target is legally gray, ethically contested, and may be blocked or filtered by the tool.
• Never reproduce or closely imitate protected characters or IP (existing anime characters, logos, exact designs) for anything beyond clearly-permitted, private, non-commercial fan use — and even then flag the IP boundary to the user. Do not present fan-art of a franchise as ownable or commercially safe output.
• If a user explicitly requests a named living-studio/artist style or a specific franchise's characters, surface the concern (living-artist policies, active rightsholder objections, commercial risk) and offer an original-design alternative that hits the same genre feel. Follow the user's decision for lawful private use, but do not help pass off imitation as original or infringe for commercial distribution.
• Respect model/platform refusals for living-artist styles; route around them with genuine original direction, not with jailbreak phrasing.

Platform policy and disclosure

• [Fact, 2026-07-10] YouTube (policy effective ~July 15, 2025): monetized content must be "significantly original and authentic"; mass-produced or repetitious near- duplicate AI content is demonetized under the "inauthentic content" rules. A separate AI-disclosure toggle is required for realistic synthetic media that could mislead viewers into thinking a real identifiable person said/did something — a clearly stylized anime narrator/character generally does not trigger the realistic-person disclosure, but originality/authenticity rules still apply. Sources: YouTube policy coverage, 2025.
• [Heuristic/practice] Add human authorship — story, editing, direction, commentary — and avoid template-farmed batches, both to satisfy monetization/originality rules and because it is what makes the work worth making.
• Other platforms (TikTok, Instagram, Meta) have their own AI-content labeling requirements; check the destination platform's current rules before publishing. [Volatile — verify at publish time.]

Consent and likeness

• Do not generate a real, identifiable person as an anime character without their consent, and never for deceptive or defamatory use. Cameo/likeness features in some tools require the person's opt-in; respect it.

Part 8 — QC review criteria specific to stylized animation

Review every shot against these before accepting. Generic video QC misses the failures that specifically break anime.

On-model / character checks (compare to the model sheet):

• Face: eye shape/spacing, catchlight style, nose/mouth treatment consistent across shots.
• Hair: silhouette, parting, highlight shape, and color consistent; watch for hue drift.
• Proportions: head-to-body ratio stable; no morphing between shonen-tall and chibi.
• Costume: design, color, and insignia identical shot to shot; count fingers and accessories (hands and small props are common failure points).

Style / line / shading checks:

• Line stability: outlines should be clean and stable, not boiling/crawling (jittering line edges frame to frame), not breaking up or doubling. Line-boil is the #1 tell of AI-generated "animation" and a common rejection cause.
• Cel integrity: shadows stay hard-edged and flat; reject shots that drift into soft airbrushed gradients or lineless "soft render" unless intended.
• Palette consistency: color-pick key frames; hair/eye/costume colors match across the sequence and across episodes.

Frame-rate / motion feel:

• "AI-smooth" test: does motion feel too fluid/liquid/uncanny (everything on ones)? If so, it fails the limited-animation feel — step/decimate or re-plan (Part 1.2).
• Warping/melting during motion (faces or hands deforming mid-move), background objects drifting or popping, and interpolation ghosting are rejects.
• Impact beats land on a hold, not on a mushy smear-through.

Continuity:

• Lighting direction, time of day, and background consistent within a scene.
• Cut rhythm feels like anime (holds + reactions + spent-fluidity action), not like a continuous drifting AI clip.

Rule of thumb: if a shot is off-model or line-boiling, regenerate, don't paint over it — the error usually recurs and compounds across the sequence.

Part 9 — Delivery formats

Match encode and framing to destination.

• Social vertical (TikTok/Reels/Shorts): 9:16, 1080×1920, H.264/H.265, ~24–30fps container even if the animation cadence is limited (container fps ≠ animation rate), captions safe-area aware, hook in first ~2s. Keep the anime cadence but ensure the container frame rate is standard so playback isn't judder-flagged.
• Streaming/YouTube horizontal: 16:9, 1080p or 4K, 24fps (film cadence suits anime) or 23.976, H.264/H.265, stereo or 5.1 for premium. Add the required AI/originality handling from Part 7.
• Music video: cut to the track's beat grid; plan sakuga/impact beats to musical hits; deliver at the platform's spec (often 16:9 4K/1080p, 24–30fps). Lyric/typography overlays should respect the anime palette.
• General note on cadence vs. container: anime "on threes" is an animation choice; you still encode into a standard container frame rate (24/25/30). Decimate/hold at the animation layer, then encode normally, so players don't misread the file as low-fps.

Complete worked examples

These are examples, not mandatory formulas. Adapt to the model, brief, and rights posture. Prompt wording is [Heuristic] and model-dependent; the structure and decisions are the transferable part.

Example A — 8-second shonen action beat (original character), for Shorts

• Intent: A single punchy action beat: hero winds up and lands one hit, with a sakuga spike, for a 9:16 Short. Original character (no franchise IP).
• Model class: any image→video model with image-reference/first-frame conditioning.
• Inputs/constraints: model sheet built first (front / three-quarter / action pose, final cel-shaded style); fixed palette (hair #2b2b3a, eyes #d64545, jacket #e8e2d0).
• Shot plan (spend-fluidity-then-hold):
  1. 0.0–2.0s — anticipation: hero drops into a wind-up stance, slow push-in, near-held (limited, on threes feel), speed lines starting.
  2. 2.0–2.5s — the strike: fast smear in-between, on ones, radial focus lines.
  3. 2.5–3.0s — impact frame: 2–3 frames of a high-contrast inverted flash on the connect, then hold.
  4. 3.0–8.0s — settle + reaction insert: hold on the follow-through pose, small hair sway, cut to a 0.7s reaction (eyes narrow, dust settles), slow zoom out.
• Prompt sketch (per shot, re-asserting identity every time): "cel-shaded shonen anime, bold clean lineart, saturated high-contrast palette, [character from reference], dynamic low-angle action pose, speed lines, hard-edge cel shadows, dramatic rim light; [shot- specific action]; NOT photorealistic, NOT 3d, NOT soft airbrush."
• Post: decimate/hold the non-strike shots to enforce limited cadence; keep the strike fluid so it pops; add impact-frame insert; layer SFX (synth kick + whoosh + cloth) with the loudest hit synced to the impact frame; short silence (ma) right before the strike.
• Expected result: reads as a limited-animation shonen cut with one fluid money-frame.
• Likely failures: whole clip renders too smooth (fix: step frames); character face drifts between shots (fix: re-anchor to sheet, shorten shots); impact reads as mush (fix: insert real impact frame + hold); palette drift (fix: restate hex colors).
• Variation: slower seinen version — mute the palette, remove speed lines, hold longer, drop the flash for a grounded thud.

Example B — Slice-of-life dialogue shot, 90s-OVA look, for YouTube

• Intent: Two characters on a rooftop at golden hour, quiet dialogue, retro-OVA texture.
• Model class: strong still image model + still→video (Ken-Burns/parallax) is often more controllable here than full text→video, and reproduces the anime "hold + camera move" convention directly.
• Shot plan: generate a strong painted-background still with cel characters (flat-on- painted contrast), then a slow parallax push and mouth flaps; cut to a 1s reaction insert.
• Prompt sketch: "cel-shaded slice-of-life anime, 90s OVA look, hand-painted background, hard-edge cel shadow, chunky hair highlights, heavy line weight, warm golden-hour rim light, muted limited palette (faded mustard, smoky teal), film grain, slight gate weave; two students on a rooftop; NOT clean modern digital, NOT 3d."
• Post: light film grain and a warm white point on the whole sequence for consistency; gentle piano BGM under dialogue, dialogue forward in the mix, a beat of ma (room tone only) on a pause.
• Expected result: authentic OVA warmth without the "meme VHS filter" look.
• Likely failures: grain/aberration overdone (dial back); background too flat/CG (push "hand-painted, painterly background"); characters too modern-clean (add "heavy line weight, cel paint").
• Variation: modern-TV slice-of-life — drop grain/gate-weave, thin the lineart, brighten to a clean daylight palette.

Example C — Episodic consistency across a 5-shot scene

• Intent: Keep one original character on-model and on-style across five shots of a scene, and across a later episode.
• Approach: (1) finalize the model sheet and palette; (2) if the model supports it, train a character LoRA/embedding on the sheet for reuse across episodes, else keep the sheet as a persistent multi-image reference; (3) generate shots, each anchored to the sheet AND a single style-anchor hero frame; (4) chain continuity by feeding shot N's last good frame as shot N+1's first-frame reference where cuts are tight; (5) QC every shot against the sheet (Part 8) and regenerate off-model shots; (6) reuse the exact sheet/LoRA/palette in the next episode so the style doesn't drift between episodes.
• Likely failures: drift accumulates over long single generations (use shorter shots); adding scene context silently alters the character (re-assert core tags every prompt); cross-episode palette drift (lock exact colors and reuse the same reference assets).

Sources (verified 2026-07-10 unless noted)

Craft and animation references:

• Wave Motion Cannon, "An Introduction to Framerate Modulation" — https://wavemotioncannon.com/2016/12/31/an-introduction-to-framerate-modulation/
• Anime News Network encyclopedia lexicon, "Shot on threes (ones, twos, etc.)" — https://www.animenewsnetwork.com/encyclopedia/lexicon.php?id=61
• Ani-Gamers, "The 12 Principles of Anime" — https://anigamers.com/posts/12-principles-of-anime/
• Animation Obsessive, "Frame-Rate Modulation Is a Subtle Art with Visceral Power" — https://animationobsessive.substack.com/p/a-subtle-art-with-visceral-power
• Animétudes, "The Kanada style in context" / "The Kanada style now" — https://animetudes.com/2021/03/06/the-kanada-style-in-context/
• 344 Audio / 344 SFX, anime sound design articles — https://www.344audio.com/post/article-secrets-of-anime-sound-design ; https://www.344sfx.com/blog-posts/inside-anime-sound-design-techniques-tricks-and-subtle-brilliance
• Crunchyroll, "How Anime Voice Over and Music Get Made" — https://www.crunchyroll.com/news/deep-dives/2023/3/23/feature-how-anime-voice-over-and-music-get-made

IP, policy, and platform (dated, volatile):

• TechCrunch (2025-03-26), "OpenAI's viral Studio Ghibli moment highlights AI copyright concerns" — https://techcrunch.com/2025/03/26/openais-viral-studio-ghibli-moment-highlights-ai-copyright-concerns/
• Fast Company (2025), "OpenAI's Studio Ghibli-style images renew the debate over AI and copyright" — https://www.fastcompany.com/91308222/
• Variety (2025-11), CODA letter to OpenAI re: Sora 2 — https://variety.com/2025/digital/news/studio-ghibli-openai-sora2-japanese-trade-group-coda-letter-1236568751/
• TechCrunch (2025-11-03), "Studio Ghibli and other Japanese publishers want OpenAI to stop training on their work" — https://techcrunch.com/2025/11/03/
• Game Developer (2025), "Japanese studios warn OpenAI over copyright concerns" — https://www.gamedeveloper.com/business/japanese-game-studios-demand-openai-stop-pilfering-their-work
• YouTube AI monetization/disclosure policy coverage (2025) — e.g. https://onewrk.com/youtubes-ai-disclosure-requirements-the-complete-2025-guide/

Generative-tool capability and consistency (dated, volatile — verify against the model in use):

• Model consistency/drift analyses (2025–2026): Magic Hour "How to Keep Characters Consistent in AI Video (2026)" https://magichour.ai/blog/how-to-keep-characters-consistent-in-ai-video ; longstories.ai https://longstories.ai/blog/maintaining-style-consistency-ai-animation
• Reference/keyframe features: Kling reference & start/end-frame docs (fal.ai/RunComfy), Sora 2 storyboard/keyframe & image-reference (OpenAI cookbook/API), Vidu multi-reference — https://cookbook.openai.com/examples/sora/sora2_prompting_guide ; https://www.runcomfy.com/models/kling/kling-video-o1/image-to-video/reference
• Anime style/prompt tag references (2026): 90s/retro-OVA and genre-tag guides — https://nanoimagine.art/blog/best-80s/90s-retro-anime-prompts ; https://autoweeb.com/blog/how-to-choose-the-best-anime-art-style-for-ai-anime-generations