DIMAU FILM STUDIOS — REPEATABLE NEURAL-RENDERING PILOT

The browser storyboard is usable without a generation GPU. The Matrix runner is
prepared, but no neural rendering has been measured in this build environment.
The user's 19 October 2026 DxDiag lists Intel UHD 630, 128 MB dedicated graphics
memory and 16 GB system RAM. The separate planned 32 GB upgrade is not GPU VRAM.
No paid compute, GPU rental, model download or licence purchase is started here.

RUN ON AN ALREADY AVAILABLE SUPPORTED LINUX NVIDIA HOST

1. Install the official Matrix-Game runtime, following its README at this reviewed
   source commit:
   https://github.com/SkyworkAI/Matrix-Game/tree/c7a88bfc039cd97695082484986e95b962390676/Matrix-Game-4
   Use a clean checkout of that commit and Matrix-Game-4/.venv. Separately download
   official resources using its pinned manifests. Source + models are Apache-2.0;
   retain upstream component and asset notices. Nothing upstream is bundled here.
2. Use a clean RGB recording from one of the THREE SUPPLIED GAME SCENES, 864x480,
   ten seconds long, without HUD/captions burned in. Keep this exact recording for
   all three runs. Arbitrary Fiji scenery is not validated by these checkpoints.
3. Download matrix-pilot.py into your working folder. Preflight:
   python3 matrix-pilot.py --preflight --matrix-dir /path/to/Matrix-Game --resources /path/to/resources --input /path/to/lowpoly.mp4 --output preflight-results
4. On a supported GPU, use the matching preset and a NEW output directory:
   python3 matrix-pilot.py --matrix-dir /path/to/Matrix-Game --resources /path/to/resources --input /path/to/lowpoly.mp4 --scene ancient-egypt --preset rtx5070ti --output pilot-results
   Supported presets: rtx5070ti, rtx5070ti-sage, rtx5080, rtx5080-sage, rtx5090,
   rtx5090-sage, realtime (A100). Sage presets require the upstream patched build.
   The plain presets avoid that optional setup. This script does not install it.

OUTPUTS AND ACCEPTANCE

- Three rendered MP4s and contact sheets, per-frame decoded hashes, process logs,
  sampled whole-device GPU usage and report.json. A run timeout defaults to 30 min.
- Exact output dimensions/frame count/FPS must match the input. Both model and
  asset manifests are verified using the pinned upstream CLI before rendering.
- Timing includes startup, loading and encoding; it is NOT the advertised warm
  model FPS. GPU usage includes every process on the selected device, not just DiT.
- Pixel equality between repeated renders is reported separately from timing.
  Compare input/output motion and geometry visually, including occlusions and
  the final frames. Passing frame counts DOES NOT prove preserved camera motion,
  low drift or production quality. No automated quality pass is invented.
- Keep the selected GPU otherwise idle. Host RAM, exact driver/kernel behaviour
  and scene transfer remain empirical tests. No consumer memory-fit guarantee.

WHAT IS DELIBERATELY NOT INTEGRATED

- RECAP: public code is interesting, but no explicit software licence was located
  during the research review. Do not commercialise that code without permission.
- FreeVideo / MiniMax H3 weights: the reviewed community licence excludes the UK
  without separate permission. FreeVideo's Apache application licence does not
  override the weight licence. No weights or executable are downloaded.
- Latent Space Jam: the storyboard exports mask-video inputs and whitespace word
  mappings; validate the exact backbone/token mapping and hardware before running
  its separate implementation. This is not a claim of end-to-end model integration.

PRIMARY REFERENCES
https://github.com/SkyworkAI/Matrix-Game
https://huggingface.co/Skywork/Matrix-Game-4.0
https://arxiv.org/abs/2610.30390
https://github.com/latent-space-jam/latent-space-jam
https://arxiv.org/abs/2610.26244
https://huggingface.co/OpenVDN/vdn-minimax-h3-edge/blob/main/LICENSE
