image → video, for one very specific little job ✶

Making a watercolor girl
breathe

Which AI tool best turns a still watercolor / sketchwash illustration into a slow, seamless-looping clip where only the world moves — cables sliding, lane-markings drifting, a soft window reflection — while she sits perfectly still and the painting never turns into a photograph?

Vertical 9:16, 5–8 seconds, loopable, very low motion. Cat-eared girl in the front seat of a self-driving taxi on the Bay Bridge before dawn, seen from behind. Animate only subtle car motion; keep her still; preserve the illustration style exactly; no face reveal, no new characters, no cuts.

The short answer

No single tool wins on everything — your job quietly asks for three hard things at once (a cinemagraph, of a watercolor, that loops). So pick by what you care about most:

push-button today

Hailuo 2.3

The only major model whose maker explicitly trains for anime, illustration & ink-wash painting — the closest thing to your sketchwash. Identity drift is rare.

≈ $0.40 / 5s · pays worldwide

keep her still

Kling (O1)

Motion Brush + Static Brush literally lets you paint which regions move and which freeze — exactly your brief. O1 does true first→last frame for the loop.

≈ $0.20–0.56 / 5s

perfectionist

ComfyUI + Wan FLF2V

Same image as first and last frame → loop is seamless by construction. An anime LoRA preserves the style by design. Run it free on your RTX A4000s.

free local · ~$0.40 on fal.ai

The recipe that matters as much as the tool

Whichever you pick, these do most of the heavy lifting for low-motion, no-drift, loopable results:

  1. 5 seconds, not 10. Drift — both geometry and style-turning-into-realism — grows with clip length.
  2. Lowest motion + a lock-off prompt. "Locked-off camera, perfectly still. Animate only: cables sliding slowly down, lane markings drifting toward viewer, soft window reflection. Everything else holds." Models add motion by default — you must suppress it.
  3. Pin the style so it can’t realism-ify: "2D watercolor illustration, flat sketchwash, hand-drawn — not photographic, no realistic texture."
  4. Loop trick: feed the same image as first AND last frame (Kling O1 · Vidu · Wan FLF2V), or add a ~1s cross-dissolve at the seam in CapCut / DaVinci.
  5. Cinemagraph it where you can (Kling Static Brush): paint only the cables & reflection as moving, freeze her + the skyline. ↞ biggest lever for “keep her still”
  6. Source image: ≥1024px, her centred with a little room around her for the motion to live in.

Every tool, scored for this job

Drops = how well each fits this brief (not a general leaderboard). style fidelity · low-motion control · seamless loop · value/ease. Search, filter, or re-sort below.

all ☁ cloud 🎨 open / local ✦ specialist ⚠ avoid
Skip for this job: Veo 3.1 & Sora 2 (realism-biased — they smooth watercolor brushwork; Sora also rejects input faces and its API sunsets Sept 2026). And wrong-category tools: Hedra (talking-head lip-sync), Higgsfield (bold cinematic camera moves), Viggle (pose/dance transfer), Genmo Mochi (obsolete). They’re in the cards above marked ⚠ where relevant.

Honest limitations

  • Researched, not hands-on tested. Compiled from official docs, the Artificial Analysis image-to-video leaderboard (Seedance 2.0 #1, Elo 1194, ~May 2026), and dated 2026 reviews — not from running each tool on your image. Style preservation on your sketchwash is the one thing only a real test settles. Budget ~$10 to prototype the same frame on the top 2–3.
  • Leaderboards are weakly predictive here. They rank general quality, not illustration-style-preservation or low-motion control — the two things that actually matter for this brief.
  • Version & price churn. "Kling 3.5" and "Wan 2.7" circulating on SEO pages are unverified; Kling 3.0 and Wan 2.2 are the confirmed current releases. A "Grok Imagine Elo 1404 #1" figure contradicts the live leaderboard — treat as stale. Re-confirm every number at point of use.
  • Moderation risk. Stylized human / anime figures occasionally trip false-positive content filters on cloud tools (Kling, Sora especially). Local / open (Wan, Hunyuan) sidesteps this entirely.
  • One conflict flagged: a March-2026 review called Hailuo 2.3 "photorealism-focused," contradicting MiniMax’s own anime/ink-wash marketing. Both are probably true (it’s broadly capable) — so test it on your art.
ai gen