Toei, Kodansha, TBS Back ¥50 Billion AI Round for Anime Tooling

Toei's AI Pivot Is Now Official, Not Experimental

The studio that defined modern anime just told investors it will hand storyboards, inbetween frames, color work, and background art to generative models — and put capital behind the promise. At the same time, Crunchyroll, the world's largest anime streamer, has reversed course on AI subtitling and dubbing, walling off the creative layer from automation. The split defines the industry's next decade: studios chasing pipeline speed, platforms protecting audience consent, and engineers choosing which side of that line to build on.

Toei Animation's FY2025 Q4 financial presentation, released May 14, 2025, listed AI integration across four production stages as a pillar of its "Technology Strategy for Improving Productivity." The same deck revealed a strategic investment in Preferred Networks Inc., the Japanese AI lab that co-developed the Scenify system Toei tested on a short film called URVAN. That experiment converted photographs of Sasebo City into anime-style backgrounds. The results convinced Toei to scale the partnership. Toei Animation's FY2025 Q4 financial presentation reported Toei joined Kodansha and broadcaster TBS in a ¥50 billion funding round for Preferred Networks, and the companies are now discussing a joint venture to co-develop animation tooling.

The production targets are specific. Toei's slides name One Piece, Dragon Ball, and Sailor Moon as franchises slated for AI-assisted seasons. The studio frames each as a productivity lever — shortening cycles on labor-intensive steps so artists can focus on creative decisions. The presentation groups these under a broader push to "establish infrastructure to create and distribute global content for both live action and animated titles" by 2026, part of a Corporate Vision 2033 targeting 50 percent overseas sales composition, as Toei's Corporate Vision 2033's data shows, and 25 to 40 billion yen in operating profit, according to Toei's Corporate Vision 2033.

Internal restructuring accompanies the tech bet. Toei upgraded its Animation Development Office into an Animation & New IP Department, signaling that AI-enabled pipelines are expected to spawn new franchises, not just accelerate existing ones. A virtual production studio, equipped with one of Japan's largest LED walls, a motion-capture stage, and a light stage for digital human creation, gives the company a physical testbed for hybrid workflows that blend live-action plates with generated assets.

The move also reflects a labor reality Toei's slides don't spell out: Japan's animator shortage has worsened for years. Entry-level animators often earn under 2.4 million yen (≈ $16,000) per year, with turnover around 25% within four years and up to 68% within eight. Inbetweening and background art are among the lowest-paid, highest-turnover roles in the pipeline. Automating them could stabilize schedules. Toei's presentation is silent on headcount impact. What it does show is a studio betting its next decade on the premise that generative AI moves from experiment to infrastructure — and that the market will accept the result.

What AI Actually Touches in the Anime Production Line

The classic anime workflow runs strictly sequential: Script → Storyboard → Key Animation → In-Between → Coloring → Compositing → Editing. Each stage once demanded specialized talent and months of calendar time for a single episode. Japan's animation industry produces over 300 titles annually (roughly 6,820 minutes of content) supported by fewer than 6,000 properly trained animators. The same labor crisis persists. That labor crisis, not creative curiosity, forced the fracture point where AI entered.

By March 2026, over 60 percent of professional animators integrate AI solutions, per aggregated studio surveys. Approximately 80% of active anime studios now utilize AI tools to some degree. Adoption is real but narrow: surveys suggest a majority of animation professionals now use AI at least occasionally, but overwhelmingly for repetitive and downstream tasks: in-between frames, background cleanup, and above all localization. What studios are largely not doing is handing authored character animation to a model, which remains commercially marginal and reputationally charged.

MAPPA and WIT Studio run AnimeFlow Pro for end-to-end production, real-time style transfer, and character consistency engines. Production I.G and Bones use AnimeForge AI for character animation, motion capture integration, and expression libraries. Kyoto Animation and A-1 Pictures deploy Quantum Animation for background generation, architectural accuracy, and lighting simulation. Smaller and mid-sized studios lean on ToonCrafter (strong keyframe interpolation), enhanced CACANi, and custom Stable Diffusion pipelines, reporting 35–55% reductions in labor costs for transitional tasks.

The technical case for automating in-betweens is simple (repetitive labor), but the artistic case against it is real — those frames shape motion quality in ways naive interpolation flattens, and skilled animators argue the 'boring' frames are where craft lives.

Human artists remain in the loop at every creative apex. Key animators (genga) still draw the frames that define performance and timing. Directors, character designers, and world-builders make the decisions AI cannot: emotional restraint, unusual camera angles, custom character designs with unique features, expressive animation with subtle emotional cues. Original concept development remains primarily human driven. Innovative animation techniques typically emerge from human experimentation rather than AI. Emotional resonance and storytelling nuance still benefit significantly from human touch. Distinctive artistic voices are difficult to develop through purely AI driven processes.

The finished clip is not the finished shot. AI artifacts (color morphing, lighting jumps between frames, warped mouths, flickering details) are normal generation output, not a sign of a bad seed, and post-production is where an AI anime crosses from 'impressive tech demo' to 'something you would actually watch.' The common finishing stack: upscale from working resolution to delivery (Topaz is a frequent choice), then clean artifacts, often frame by frame, in a compositor or editor (DaVinci Resolve is standard for grading and assembly).

New hybrid roles have emerged that specifically focus on human-AI collaboration. Technical directors increasingly specialize in AI system customization. Quality control positions now emphasize maintaining artistic consistency across AI generated and human created elements. Key animators focus more on creative direction and distinctive moments. Studio representatives have emphasized that their primary key animators still work largely by hand, with AI assisting in maintaining quality across production teams. The most successful implementations treat AI as a style enhancement tool rather than a style generator, preserving the artistic diversity that defines the medium.

Long-form continuity is unsolved; holding characters, props, locations, and emotional performance stable across a full episode still needs heavy human supervision, and the consistency methods above reduce drift rather than eliminate it. Acting precision is weak: models render motion, but the specific timing and restraint that make a performance read are hard to direct and easy to lose. Revision control is clumsy: asking for a small, exact change to a generated shot often means re-rolling and hoping, not editing. And cost and time are real: per-second model pricing plus the re-rolls a style-critical, drift-prone medium demands means a full episode is many generations, not one.

The Backlash: Studios Apologize and Fans Demand Consent

The flashpoint arrived April 4, 2026. Twitter user ukloim posted screenshots from the opening sequence of Ascendance of a Bookworm season four — generative AI artifacts smeared across background art. Wit Studio, the house that built Attack on Titan and co-created Spy × Family, had let AI-generated backgrounds into a prime-time broadcast. Fans recognized the texture immediately. The backlash spread across X in hours, not days.

Six days later, Wit Studio folded. Their April 10 statement didn't dodge: "While we at our company are always interested in and closely monitor new technologies related to video production, we have, in principle, not permitted the use of generative AI in the video production of our works, including this one. Despite this, the current situation has occurred solely due to shortcomings in our production management and inspection systems, and we take full responsibility for this series of events." They apologized to fans, the original author, and "all other related parties." The opening sequence would be replaced with hand-drawn art starting episode two, slated for Crunchyroll on April 11. The original upload vanished from Crunchyroll's YouTube channel. Wit Studio explicitly cleared NAM HAI ART, the background studio on the project; the failure was internal quality control.

Netflix Japan took a different line. They framed the AI usage as "an experimental initiative to support the labor-short anime industry," noting image generation handled background artwork for "all cuts in a 3-minute video." The credit roll had listed background design as "AI (+ Human)," reducing the human artist to a parenthetical.

The Wit incident wasn't isolated. Director and animator Terumi Nishii (Jujutsu Kaisen, Death Note) posted on Bluesky that several animation studios are already "using AI to generate rough drafts for direction corrections," making it nearly impossible to detect when outsourced work arrives pre-contaminated. Nishii's observation matches what production insiders have whispered for months: AI drafts enter the pipeline at the layout stage, then get cleaned up by human hands, leaving no paper trail.

Creators are exiting. Three manga artists deactivated social media accounts in recent months after coordinated harassment campaigns tied to AI disputes. One Japanese mangaka watched annual earnings collapse from roughly $1,300 to $130 after AI-generated manga flooded digital marketplaces. A Nikkei survey found 22 percent of creators reporting income declines linked to AI; 6 percent had already left or were preparing to leave the profession.

Fan communities aren't just complaining — they're codifying norms. Academic interviews with 20 creative workers across visual art, writing, and programming reveal a shared framework: consent, credit, compensation. The "three Cs" have governed fan production for decades; now they're being applied to generative models trained on artist output without permission. The tension is structural. Studios want pipeline speed. Audiences want proof of human labor. The middle ground — where AI assists but doesn't replace, where credits reflect reality — keeps shrinking.

Wit Studio promised revised production guidelines. The next test comes when a studio doesn't get caught.

Crunchyroll's Counter-Move: Humans for Subtitles, AI for Discovery

The reversal took thirteen months. In February 2024, Crunchyroll president Rahul Purini told The Verge the company was "very focused on testing" AI for subtitling and closed captioning: speech-to-text pipelines that could shrink the gap between Japanese broadcast and global launch. Dubbing was on the table too, though he conceded the tech wasn't there yet for adapted scripts, humor, and lip-sync. By April 2025, the line had hardened. Purini told Forbes and Radio Times that Crunchyroll would not use AI in any creative lane: not subtitles, not dubs, not voice performances. "We consider them to be creators because they are contributing to the story and plot with their voice," he said. The platform would keep AI strictly to back-end systems (recommendations, personalization, content discoverability), tools that improve the viewer experience without touching the work itself.

Fan pressure drove the turn. Subtitle quality had slipped visibly across the service. When Necronomico and the Cosmic Horror Show debuted, German viewers found the opening line "ChatGPT said:" baked into the subs, a third-party vendor's fingerprint left in production. English tracks drew complaints for sloppy phrasing. Crunchyroll traced the German incident to a contractor, called it a direct violation of their agreement, and opened an investigation. Full findings never published, but the message to vendors landed: the platform owns the quality floor. Translators and voice actors, the people who carry cultural nuance across oceans, breathed easier. One flat AI line can flatten a joke or strip emotion from a quiet moment. Purini's clarity protects that craft.

The split inside Sony is deliberate. Aniplex, A-1 Pictures, and CloverWorks (all Sony Group siblings) continue building AI tools aimed at production tasks like lip-sync automation. Crunchyroll sits apart. The streamer invests heavily in anime committees and global reach, yet it chooses to keep the creative core human. Purini sees AI value elsewhere: recommendation engines that surface the next series, personalization that helps a viewer in Brazil find a show made in Tokyo. Those systems get the green light because they improve discovery without touching story or performance.

The business logic tracks. Crunchyroll's expansion into India, Mexico, and Brazil surfaces a consistent demand: fans want their own stories told in an anime style with Japanese production values. The company sources Indian or Korean-American narratives, takes them to Japanese creators, and produces globally — The Beginning After the End adapted a Korean-American web novel on a South Korean platform with Japanese animation for worldwide release. Digital manga joined the catalog in 2025. A new division, Crunchyroll Storyworks, launched in October 2026 under Sony veteran Tohru Iokibe to fuel anime and live-action adaptations. Sony Pictures Entertainment chairman Ravi Ahuja announced it at the company's inaugural thought-leadership event. The strategy bets on creator authenticity as a differentiator in a market where Amazon Studios is piloting AI dubbing and startups pump out AI-generated micro-dramas by the thousands.

Purini admits the stance can shift. For now, the line holds: human hands on the creative work, machine learning on the discovery layer. The engineer building a recommendation model at Crunchyroll ships code that serves the viewer. The engineer building a lip-sync tool at CloverWorks ships code that serves the studio. Same parent company. Different consent boundaries.

The Engineer's Fork: Building Animation Tools in a Divided Market

The split between Toei's production-line automation and Crunchyroll's consumer-facing retreat has created two distinct engineering tracks, and the technical requirements for each barely overlap. Video model labs, studio vendors, and AI-native startups are hiring for different problems, on different timelines, with different risk profiles. An engineer choosing among them is effectively choosing which side of the consent boundary they want to build on.

At the model labs, the divergence is already visible in product philosophy. Luma AI's Dream Machine, built around the Ray3 stack and its January 2026 Ray3.14 update, targets professional pipelines: native 4K HDR output, ACES EXR export, multi-keyframe sequencing up to 16 frames, and a reasoning system that evaluates its own generations for physical plausibility before finalizing. Ray3.14 runs four times faster and three times cheaper in credit terms than the launch version while preserving the HDR pipeline. Luma Agents stay with a project from brief to final delivery, handling revisions across video, images, and audio while maintaining creative context. The platform integrates with Adobe Firefly and Amazon Bedrock, syncing content to Premiere Pro, infrastructure built for studios that need footage to drop into existing timelines. Luma generates 5-second clips in roughly 22 seconds.

Pika took a different fork. Its September 2026 relaunch of pika.art as a multi-app studio routes each request across Seedance 2.5, Pika 2.5, MiniMax H3, Wan 3.0, Google Omni Flash, GPT Image 2.5, Nano Banana 2, Grok Imagine, and Pika's own audio models, featuring automatic model routing with manual override. The Pikaffects suite (melt, inflate, explode, cakeify, Pikatwists) drives engagement on TikTok and Reels; Pikaframes generates transitions between user-defined start and end images with frame-level precision. Pika generates 3-second clips in 12 seconds. Its $10/month API Club offers wholesale access to 100+ models at up to 88% savings versus other aggregators: MiniMax H3 T2V 2K at $0.13/second, Seedream 5 Pro 2K at $0.09/image. Four foundation audio models (Music, Soundtrack, SFX, Speech) run up to 20× cheaper than comparable alternatives; Pika Speech is 9× more cost-efficient than ElevenLabs v3. The MCP toolkit lets agents like Claude, Cursor, and ChatGPT generate media directly. Pika exports optimize for immediate posting: aspect ratios match TikTok and Instagram specifications, file sizes suit direct upload. No enterprise platform integrations comparable to Luma's Bedrock and Adobe deployments exist yet.

Studio vendors sit on a third track. Toei's investment in Preferred Networks signals a different procurement model: co-development toward a joint venture, targeting storyboard creation, coloring, automatic correction of genga (key frames), and photograph-to-anime-background conversion. The requirement isn't a general-purpose video model — it's a tool that fits the anime production line's specific stages, where human key animators still draw the genga and AI handles inbetweening, color correction, and background generation from reference photos. Engineers here build for deterministic output, version control that animators trust, and integration with existing studio software (CLIP STUDIO PAINT, RETAS, custom pipeline tools). The consent boundary is internal: the studio decides what enters the pipeline. The technical bar is reliability at production scale, not viral clip speed.

Crunchyroll's retreat from AI subtitling but doubling down on AI discovery (recommendations, personalization, content tagging) adds a fourth track: backend ML for a walled garden where the model never touches the screen; it only decides what the subscriber sees next.

No single track is winning. The market is splitting, and the engineers who understand which constraints they're optimizing for — pipeline compatibility, social velocity, studio determinism, or discovery relevance — will be the ones who stay employed when the hype cycle corrects. The next hiring wave won't ask "do you know diffusion models." It will ask "which production line have you shipped into."

Working in frontier tech? Zero G Talent tracks the openings: see every open ASML role, browse frontier tech jobs, openings at Stripe and Waymo, and the people building the field.