Kikitora discloses $19.7M round for offline AI motion capture

Why it matters

Kikitora raised nearly $20 million while still at waitlist stage, backing a founder-led attempt to make private, hardware-free motion capture usable across animation and robotics.

Astrid Wilde, a technical artist and founder, has raised $19.7 million for Kikitora, a San Francisco company building AI software that converts ordinary video into 3D performance data.

The financing appeared in a Form D filed with the SEC on August 11th. Kikitora sold the full $19,743,837 securities offering to 43 accredited investors, with the first sale recorded on July 22nd. The filing names Wilde as chief executive and director. It does not identify the investors, a lead backer or Kikitora's valuation.

Kikitora incorporated in Delaware in 2025, making the offering substantial for a company that is roughly a year old and whose public website still directs prospective users to a waitlist. The SEC filing lists Kikitora under "Other Technology" and declines to provide a revenue range. That leaves product performance, customer adoption and commercial terms outside the public record even as the financing gives Wilde considerable room to build.

Wilde's background fits the problem. Her technical-art portfolio says she studied Games and Interactive Media at Boise State University, led a 10-person art team while building a networked character creator and developed a real-time procedural water system for the surfing game YouRiding. Her work sits between software engineering and the production artists who have to make animation tools usable inside demanding pipelines.

That experience matters because motion capture remains partly a workflow problem. A system can produce convincing pose estimates in a demo and still fail a studio when hands disappear behind a body, two performers overlap, clothing moves independently or a camera changes focus. Kikitora is trying to cover those cases with one software stack rather than requiring artists to assemble separate tools.

Turning existing footage into animation data

Kikitora says its software accepts RGB or grayscale footage from smartphones, webcams, professional cameras and existing video. It claims to recover full-body movement, facial expressions, finger-level hand motion, camera information, multiple performers, props and deformable objects such as cloth.

Kikitora also advertises real-time streaming into Maya, Blender, Unity, MotionBuilder and Unreal. The pitch is aimed at animators, game developers and virtual-production teams that want performance capture without dedicated suits, markers or calibrated stages.

The use of existing footage widens the proposition. Kikitora says films, sports recordings and internet videos can become starting material for 3D animation. That could make motion libraries far larger and cheaper to assemble, though Kikitora has yet to publish benchmarks showing how its system performs across camera angles, occlusion, lighting conditions or heavily compressed video.

Local processing is Kikitora's clearest commercial distinction. Kikitora says footage can remain on the customer's device and that the system works offline. Studios handling unreleased characters, actor performances or licensed footage have reasons to avoid uploading those assets to a third-party service. Robotics developers working with human demonstration data face similar controls around proprietary environments and recorded participants.

A technical foundation with disclosed limits

The clearest public evidence of Kikitora's computer-vision work is CURDIE3, an Apache 2.0-licensed synthetic pose-estimation dataset co-authored by Wilde and six other contributors.

The 919 GB repository contains 1,408,410 rendered images covering 100 actor models, 1,400 animations and three environments. Its labels include body, hand and foot keypoints, camera parameters, depth maps and instance-segmentation masks. Those are directly relevant inputs for systems attempting to reconstruct people and camera movement from two-dimensional images.

The dataset card also documents meaningful constraints. Its actors have neutral faces, move independently and can intersect with one another or the environment. Feet may penetrate floors or float, and the camera motions use a limited set of constant movement patterns. The authors say CURDIE3 is unsuitable for facial-expression recognition, social-interaction modeling and applications requiring physically plausible motion.

Those disclosures make CURDIE3 useful evidence of Kikitora's engineering approach while setting a boundary around what it proves. The dataset alone cannot validate Kikitora's broader claims around facial capture, coordinated performers, cloth or production-quality output. Kikitora will have to demonstrate those capabilities in the product.

An established market with room for a privacy-first tool

Kikitora is entering a field with several commercial systems already turning video into animation. DeepMotion offers browser-based video-to-3D animation, Move AI exposes motion capture through an API, Autodesk Flow Studio generates CG scenes from video, and Rokoko sells software alongside dedicated capture hardware.

Wilde's bet is that a broader capture system running locally can earn a place beside those products. Kikitora's advertised scope spans the performer, face, fingers, camera and surrounding objects, all from footage recorded without specialized gear. Delivering that combination consistently would reduce both capture costs and the cleanup work that follows.

The $19.7 million round gives Kikitora a financing base well ahead of its public product availability. It also raises the standard Wilde now has to meet. Customers will need evidence that Kikitora can preserve subtle performances, handle difficult scenes and export dependable data into the tools artists already use.

For Wilde, Kikitora is a direct extension of years spent building tools for artists rather than a departure into an unfamiliar market. Investors have funded that domain experience before Kikitora has shown a generally available product. The next test is whether Wilde can turn technical-art fluency and computer-vision research into software that survives real production footage.