# Asset provenance and display behavior

## Saved poses

Repository: https://github.com/NVIDIA/synthda

Pinned revision: `d989c0148c72537b507a0e3e6b80f46685a6d3b0`

Samples under `components/dataset/data_manipulation/`:

| Viewer example | Repository folder | Source A | Source B |
| --- | --- | --- | --- |
| Real + generated 01 | `test_video_1` | `output_keypoints_3d.npy` | `gen_motion_00_L108_00_a_flip_extended.npy` |
| Real + generated 02 | `test_video_2` | `output_keypoints_3d.npy` | `gen_motion_00_L072_00_a_flip_extended.npy` |
| Two real motions | `test_video_1_test_video_2` | `output_keypoints_3d_real2.npy` | `output_keypoints_3d_real1_extended.npy` |

The real-real source ordering follows the executed branch in
`optimisation_both_real.py`, where the longer second sequence becomes A.
All 27 saved arrays were checked against the source implementation:

`P_opt = P_A + 0.5 * (wA * P_A - wB * P_B)`

This is the algebraic result of the distance/direction operations in
`compute_P_opt`. It is **not** the ordinary convex blend
`wA * P_A + wB * P_B`. We therefore show weights as saved optimizer parameters
and do not imply that 0 and 1 smoothly recover the source endpoints.

Coordinates are serialized to five decimal places for delivery. The stored arrays are unchanged; no floor snapping or new model generation is applied.
The browser defaults to a display-only symmetric Gaussian temporal filter (nine
frames, sigma 1.8 frames) and interpolates between frames during playback. The
Original option bypasses the filter. Downloads always use the original arrays.
The articulated mannequin is a procedural body representation driven by source A
joints, not a fitted SMPL surface, identity reconstruction or augmented result.
Foot joints duplicated in the source mapping remain duplicated.

Source videos are the original `test_video_1.mp4` and `test_video_2.mp4` files.
Each is matched to its same-folder extraction. RGB time is mapped to normalized
pose progress, not verified action-event alignment. Generated and real motions
may perform the same action at different relative times.

The stage uses a display-only orthographic projection of all three joint
coordinates. Scale is shared across comparison lanes and derived from the
entire example to avoid per-frame resizing. The grid is decorative spatial
reference; root-relative data does not reconstruct the actor's world-space
height or physical ground contact. Default playback is 25 fps for presentation,
not a claim about the original acquisition frame rate.

## Original Space archive

Repository: https://huggingface.co/spaces/nvidia/synthda-demo

Pinned revision: `dc65e86fc69dcb339b90d68053666bc6d3604e5c`

The three folders `synthda_falling_realreal`, `synthda_demo_fall_2`, and
`synthda_demo_fall_3` each contribute eleven unmodified MP4s. The previous
app's real/generated labels conflict with its collection titles and folder
names. The archive therefore uses neutral collection and clip identifiers.
The archive clips are not asserted to correspond to the playground pose arrays.
`assets/media-provenance.json` lists exact URLs, SHA-256 hashes and byte sizes.

## New motion studies

`compare-study.mp4` and `spectrum-study.mp4` were rendered from the first public
pose example. They are 1280×720 skeleton studies, not SMPL surface renders or
new model outputs. They use synchronized frame indices and preview timing.

`assets/social-preview.png` is the actual earlier interactive-preview capture.

## Assets still required for the expanded showcase

- Additional action classes with original clips and corresponding pose outputs.
- Mesh sequences aligned to the exact poses, frames and coordinate convention.
- Verified recognition results with dataset, split, metric, baseline, augmented
  value, model configuration, and a source or evaluation artifact.

These are intentionally absent from the initial manifest. No body surface has
been inferred from skeleton coordinates, and no recognition uplift is claimed.

## Smoothing check

For real-plus-generated example 2, variant wA 0.5, the coordinate-wise second-
difference RMS over frames 76–130 decreased from 0.02153147 to 0.00113265
in the stored coordinate units. This is a numerical smoothness check, not
a measured recognition or motion-quality score. The symmetric filter may soften
fast motion details; use Original to inspect the unfiltered extraction.