Comic

Comic is our family of motion perception and quantification models.

Try Comic

Comic captures human motion from monocular video, estimating not just depth, pose, scale and camera position, but the physics beneath the motion — gravity, ground contact, and the forces that yield position, velocity, and acceleration — to deliver industry-leading performance.

Comic captures production-grade motion directly from video, at a fraction of legacy cost — collapsing the tradeoff between quality and scale, and making it viable to generate motion data in the volumes physical AI requires.

Our customers use it for gameplay and narrative animation, robotic motion authoring, expert trajectory data, AI model training data, neural rendering, sports performance, and clinical movement analysis.

Models

Comic 4

Multi-subject, faces, scale, and camera capture in the highest fidelity

Try Comic 4

Comic 3

Fast, single-subject body performance capture.

Try Comic 3

Comic Realtime

Realtime motion capture & retargeting for streaming or recording

The fastest, most accurate, feature-rich video performance capture model.

Capabilities

Multi-subject

Accurate subject scale

Faces

Hands, wrists, & fingers

Camera intrinsics & extrinsics

Foot/ground contact

Accurate world space, root translation, and height

Objects

Environments

Performance

20406080100

Historical EMDB benchmark · September 4, 2026 · higher normalized score is better.

Historical EMDB benchmark · 39 clips · 62,707 frames. Predates later MHR audit corrections.

Historical benchmark methodology

Successful clips: Comic 4 39, Comic 3 38, Comic Realtime 39. Comic 4 uses the saved anchor2 skeleton benchmark; Comic Realtime uses saved moro_static static-camera world placement. Error axes use 100 × clip(1 − value/cap, 0, 1) with the original generator default cap of 1.25 × the worst selected model. Speed uses a logarithmic 0.01×–1× realtime range.

Comic model performance comparison

Historical EMDB benchmark data. Raw error measures are lower-is-better; their supplied normalized scores invert errors so higher is better. Results predate later MHR audit corrections.

Historical EMDB raw benchmark values and normalized scores by metric
ModelPose accuracyMotion fidelityWorld accuracyTrajectoryRoot driftSmoothnessFoot contactSpeed
Comic 451.58 mm; normalized score 22.2/1003.22 m/s²; normalized score 87.6/100118.03 mm; normalized score 81.8/100264.74 mm; normalized score 81.2/1003.19 % of path; normalized score 79.9/10015.89 (10 m/s³); normalized score 75.1/1004.41 mm; normalized score 95.4/1000.055×; normalized score 36.8/100
Comic Realtime51.68 mm; normalized score 22.1/10020.78 m/s²; normalized score 20.0/100517.59 mm; normalized score 20.0/1001128.56 mm; normalized score 20.0/10012.70 % of path; normalized score 20.0/10050.97 (10 m/s³); normalized score 20.0/10077.07 mm; normalized score 20.0/1000.285×; normalized score 72.7/100
Comic 353.04 mm; normalized score 20.0/1005.15 m/s²; normalized score 80.2/100133.21 mm; normalized score 79.4/100306.45 mm; normalized score 78.3/1003.41 % of path; normalized score 78.5/10024.71 (10 m/s³); normalized score 61.2/1003.11 mm; normalized score 96.8/1000.056×; normalized score 37.4/100

See Comic in action

Try Comic thru the Cartwheel MCP or in the Performance Capture app.