Skip to content
Bud's Wiki

CYANPUPPETS 1B real-time motion model (camera → UE/Unity/Blender)

TL;DR

Section titled “TL;DR”

A demo tweet from CYANPUPPETS (an NVIDIA Inception studio, 2023 NVIDIA Showcase) announcing a 1-billion-parameter real-time motion model that ingests a 1080P webcam feed or uploaded video and outputs character motion streamed live into Unreal Engine, Unity, and Blender. The model runs at real-time rates on 8 GB of VRAM. No paper, weights, or method writeup is linked — this is a product-demo tweet.

Key claims

Section titled “Key claims”
  • The model has ~1B parameters and is described by the author as an “AI real-time motion model” built on computer vision and kinematics [tweet body].
  • Input is a 1080P camera feed or an uploaded video; output streams to Unreal Engine, Unity, and Blender [tweet body].
  • Real-time processing requires 8 GB of VRAM [tweet body].

Method

Section titled “Method”

Nothing technical is disclosed in the tweet. The author bio describes the system as a “motion model built on artificial intelligence, computer vision, and kinematics,” which is consistent with a markerless monocular motion-capture pipeline — most likely video → 2D keypoints → 3D pose lifting → retargeting to a rigged character — but no architecture, training data, or evaluation is given. The 1B-parameter / 8 GB-VRAM / 1080p-realtime envelope is what’s claimed; the demo video showing live character animation in DCC software is the evidence offered.

Results

Section titled “Results”

No quantitative results. The artifact is a ~3-second demo clip embedded in the tweet, viewed ~35K times as of filing.

Why it’s interesting

Section titled “Why it’s interesting”

A concrete data point on the engineering envelope for production-grade markerless mocap: 1080p real-time at 8 GB VRAM with a 1B-parameter model, streamed into the three major DCC engines. Worth tracking as a deployment reference point alongside the research literature on monocular pose estimation (e.g. FMPose3D: monocular 3D pose estimation via flow matching) — academic lifters publish FID-style accuracy on H36M / 3DPW but rarely report this kind of latency × VRAM × engine-integration profile. If/when CYANPUPPETS ships a tech report or model card, that’s the artifact to file; this page is a placeholder for the demo.

See also

Section titled “See also”