Reconstruct with AI - new button for body captures that rebuilds the animation with NVIDIA MotionBricks: key poses come from the capture and the motion between them is generated
Before / After comparison - the capture and the AI version play side by side, labelled BEFORE and AFTER, in sync and turning together when you orbit either one
Export both - Export writes the capture and the AI version together, the AI version as <name>_rebuilt.fbx
Model download - the MotionBricks model (about 730 MB) downloads once on first use, with a progress bar; it is shared with AI Animator, so an existing install is reused
Progress - the animation area shows the loader while the model downloads and while reconstructing, with a cancel button
Looks like the video - the AI version follows the performance closely: hips, spine and limbs are pulled back wherever the AI drifts from the capture
Smoother motion - the AI version is about half as jerky as the capture on the test clips
Kick height - kicks keep their full height instead of coming out low
Planted feet - feet the performer keeps planted stay planted in the AI version; feet that move in the video still move
Wrists - wrists keep the capture's bend, so they no longer wobble or spin
Hands behind the body - a hand that passes behind the body in the video (floss, chambered fists) now goes behind it instead of hitting the torso
Reset camera button - centred on the character dropdown
Sliding feet - the AI version's feet no longer slide while planted
Joins - removed the hitches where generated pieces of motion met
Humanoid Retargeter required - Reconstruct with AI needs the Humanoid Retargeter library installed in the project
Body captures only - Reconstruct with AI is not available for First Person hand captures
Capture-only parts - fingers, wrists, neck and head always come from the capture
Very quick moves - moves quicker than about half a second (such as fast repeated kicks) come mostly from the capture
Automatic depth detection - checks every video for phone depth data and uses it automatically; videos without depth work as before
Full LiDAR - Record3D recordings (RGBD video or .r3d) are labelled FULL LIDAR DETECTED and the body is placed at the distance the sensor measured
Measured hand distance - clearly visible hands in LiDAR recordings take the distance the sensor measured
Phone camera path - moving .r3d recordings follow the camera path the phone tracked instead of an estimate
Partial depth - iPhone Cinematic mode videos are labelled PARTIAL DEPTH DETECTED and their depth corrects hands in front of or behind the body
Depth border - the source preview pulses green for full LiDAR and amber for partial depth
LiDAR view - new button beside the source preview shows a colour depth view (red near, blue far); disabled when no depth is found
.r3d files - Record3D .r3d files can be uploaded, dragged in or picked with Upload video…
Body distance - on LiDAR recordings the body's distance from the camera went from up to ~20 cm off to about 1 cm
HEVC videos - H.265 videos preview and process without the Microsoft Store HEVC extension
Phone upload - depth tracks are kept; the page explains how to re-send a Cinematic video whose depth was stripped
Hand placement - hands match where the video shows them, including above the head, crossed and against the chest
Wrist direction - palms face the right way, taken from the hand tracker
Joint limits - elbows and knees only bend the way real joints do; fingers stay within natural limits
Foot contacts - planted feet stay put across the whole clip
.mov preview - iPhone .mov videos no longer show "cannot preview" or play frames out of order
Decoded timestamps differ - fixed the capture error that appeared after a recent Windows update
Arms through the body - arms no longer pass through the torso
Floating hands - hands on the floor stay planted, for example when getting up from sitting
Cinematic depth from Photos - picking a Cinematic video from Photos on the phone upload page strips its depth; choose it from Files ("Choose File") or copy it with Keep Originals
Depth changes distance only - poses, joint angles and fingers still come from the video
Upload screen - dark drop box with a green Choose Video button, like Humanoid Rigger'saeafa4c
Processing loader - the animation panel shows a spinner and the current step while a video processes0b37391
Download progress - a progress bar with the file name and how much is left to download0b37391
Reinstall worker - a button in Advanced that deletes and re-downloads the inference worker7b117a8
Prebuilt worker - installs from s&box download the inference worker automaticallye3cf2c7
Workspace - redesigned with cards for the source video, animation and playback, and a status lighte38bb93
Advanced settings - grouped into Capture, mode options, Cleanup, Contact review and Capture detailse38bb93
Preview camera - renamed Head cam / Orbit and shown only in First Person, so it no longer clashes with the First/Third Person choicee38bb93
Preview framing - the camera holds steady instead of re-framing every frame5c75b26
Smoothing - speed-adaptive, so fast moves stay sharp and slow ones stay smoothb986e92
Camera tilt - captures are levelled with gravityf5f9cc3
Jumps and flips - flights keep their real height, with smoother takeoffs and landings0ea51b7
Sitting - hips sit on the floor, and seated hips and feet stay puta3a8d25
Spine - straight spines with no mid-back hunch or crease at the base978fc67
Travel - tumbling and rolling moves stay on their linea82ad09
Wrist direction - hands face the way the video shows, using the finger model where it agrees with the body5665575
Viewmodel arms - a body capture is shown on the full character when arms are selected9b8c1d3
Damaged videos - captured up to where they stop playingaf37444
First download - Third Person needs about 9 GB of models on the first capture (First Person about 2.6 GB)
Moving cameras - clips filmed with a moving camera can still show some foot slide
Two editors - starting a first capture in two editors at once can make the second download fail; press Retry