# Spatial photo demonstration — sources and method

Retrieved 5 October 2026. Photographs are used under the Pexels licence: https://www.pexels.com/license/ . The source files are resized for this demonstration. Depth displacement modifies their presentation. No brand partnership or product offer is implied. Photographers are credited in the studio.

| Photo | Photographer | Source |
|---|---|---|
| Trainers | Ox Street | https://www.pexels.com/photo/photo-of-white-sneakers-6050912/ |
| Headphones | Guillaume Meurice | https://www.pexels.com/photo/black-headphones-6587325/ |
| Interior | Skylar Kang | https://www.pexels.com/photo/living-room-interior-at-home-in-sunlight-6044718/ |

## Depth model

Depth Anything V2 Small, 24.8M parameters, via the authors' Hugging Face checkpoint. Model: https://huggingface.co/depth-anything/Depth-Anything-V2-Small-hf . Official repository and licensing: https://github.com/DepthAnything/Depth-Anything-V2 . The Small checkpoint is Apache-2.0; larger variants have different licences and were not used. No model weights or AI packages are included in the site. The depth maps are model outputs.

Citation: Lihe Yang, Bingyi Kang, Zilong Huang, Zhen Zhao, Xiaogang Xu, Jiashi Feng and Hengshuang Zhao. Depth Anything V2, NeurIPS 2024.

## Method and limits

Each photograph is paired with an offline estimated relative-depth PNG. Predictions are normalised to the 1st/99th percentiles, lightly smoothed, and mapped to a shallow textured grid. 2D sets every vertex to the same plane; Spatial varies the depth by up to 0.42m behind that plane. A perspective adjustment preserves the source view at the nominal 4m viewpoint. Both eye cameras render the same geometry naturally in WebXR, creating binocular disparity and modest head-motion parallax. The image texture is unchanged between modes.

This is a prototype for limited viewing angles. Mesh stretching and imperfect depth boundaries may become visible, especially from large sideways movements. It does not fill occlusions, reconstruct the back of products, reproduce Meta's proprietary conversion, or provide metric product dimensions. A merchant pilot needs image-by-image quality assessment and a flat-image fallback. Video conversion is outside this milestone. Physical Quest comfort and visual quality need user validation.

## Consumer film in the Common Room

Source: the user's existing Loom_Consumer_Experience_Inspirational.mp4 (48 seconds, current supplied version). Existing visuals and soundtrack are retained; no new retailer footage is introduced. Transcoded for playback with FFmpeg. The depth atlas was generated offline with the same Apache 2.0 Depth Anything V2 Small model, sampled at 4fps and interpolated in a 96 × 54-segment mesh. Relative depth is an artistic estimate. Light temporal smoothing and caption-region masks reduce artefacts but do not guarantee stable motion, accurate geometry, complete caption separation or comfort. This is a prototype 2.5D film presentation, not captured stereoscopic footage or a traversable world.

## Investor’s Hall presentation

The 25 rendered JPEG slides derive from Brad’s existing Investor_Deck.pptx, version 1 dated 3 October 2026. The original editable deck was not modified. Web rendering reduces text size in memory to keep columns legible without overlaps. The presentation is an early concept/funding draft and contains planning assumptions rather than validated traction. The supplied concept film remains the original 2D video; spatial playback controls are parked.