Most AI video tools today are trying to be everything to everyone — photoreal cinematic clips, fashion ads, real-estate walkthroughs, GIF memes. They're all converging on the same general-purpose render endpoint, just with different chrome around it.
We picked a different angle: anime and animation creators. Not because the market is small (it isn't — anime is a $25B industry and growing) but because their tools don't exist yet. Hand-drawn animators get the worst of every AI tool: photoreal drift mid-motion, butter-smooth 30fps instead of held-on-twos cadence, and zero respect for the visual vocabulary of the medium (no impact frames, no speed lines, no manga panels).
What we shipped
Over the last few weeks KOANimation Studios went from "generic AI playground" to a tool actually shaped for animation:
- 30+ AI models behind a unified interface — Sora, Veo, Kling, Wan, Seedance, Flux, Ideogram, Midjourney. Pick the right model per shot, not the right tool per task.
- Character LoRAs — train an image LoRA on 4-20 references, get a recurring character across every render. Video LoRAs too, for Wan-based motion consistency.
- Lock to 2D — a procedural cel-shading post-process that flattens textures, boosts saturation, and emphasises outlines. Snaps photoreal output back toward hand-drawn aesthetic.
- Anime VFX library — manga panel borders, impact frame flashes, radial speed lines, sakura drift. Procedural ffmpeg, zero AI, 3 credits each.
- Cadence retiming — convert any clip to anime on-twos (12fps held) or smooth 60fps interpolated motion. Single click.
- Character reference sheets — turntable, expression sheet, pose sheet. Batch-generates 5/9/8 images of your character at once.
- AI Continuation — extend any video clip by feeding its last frame back into a different i2v model. Picker shows IP policy per model so you know what's strict (Veo, Sora) vs open (Kling, Wan, Seedance) before you burn credits on a refusal.
- Tween A → B — upload two key frames + motion prompt → generated video between them. The post-ToonCrafter workflow.
What's hard about this niche
The hardest part isn't building features — it's the unglamorous bits AI tutorials skip:
- Routing across providers. fal.ai, muapi, direct OpenAI and Anthropic — every model has different input schemas, different field names, different async patterns. We built a provider-aware submit layer that the UI consumes uniformly so we can swap any model out without rewriting the frontend.
- Credit accounting. Each fal model has its own per-megapixel or per-second pricing. We deduct via Postgres RPC before submit, refund automatically if the model fails. Every penny is tracked in a ledger so we can spot abuse and reconcile against Stripe at the end of each month.
- IP policy at the model level. Veo and Sora refuse copyrighted characters; Kling, Wan, Seedance shrug. We tag every video model with a policy badge and warn users before they submit a prompt that's clearly going to bounce.
What we learned
Three things, in order of importance:
- Pick a niche. "AI video for anime creators" is a completely different positioning than "AI video tool." It defines what features matter, who you market to, and which complaints to take seriously. Generic AI video tools end up competing with Higgsfield and Pika on price; niche tools build moats.
- Procedural beats AI when it works. Speed lines, manga panels, on-twos cadence — these don't need a new model. They need ffmpeg and a picker UI. Procedural effects are instant, cheap, and deterministic. AI is the long tail.
- Ship the unsexy stuff. Cookie banner. GDPR delete endpoint. TOS that mentions AI specifically. Refund policy. These aren't features users brag about — but skipping them is what turns your first 10 customers into 10 chargebacks.
Try it
Free trial: 50 credits, no card required. Sign up at koanimationstudios.com and jump straight to the studio at /studio.
We’d love your feedback — contact us if anything breaks, anything works unexpectedly well, or you have an animation use case we haven’t thought of.