Files
saphid--frame-control/docs/vr-video.md
T
saphidandClaude Opus 5.5 f81f70ed5d Fix media stop, upload cleanup and review findings
- Stop is a no-op when the collected player unit is already gone
  (raw systemctl stop exits 5 on the Frame; verified 2026-09-29).
- Surface systemd-run stderr when the player can't start.
- Keep the copy error if the cleanup ssh also fails; reject upload
  names that the play path can never accept.
- Allow 60 s for play (ffprobe 30 s + systemd-run 15 s remote).
- Docs: four-hour cap is unconditional; no delete action yet; fix a
  garbled timing sentence.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
2026-09-29 09:22:22 +10:00

149 lines
8.0 KiB
Markdown
Raw Blame History

This file contains ambiguous Unicode characters
This file contains Unicode characters that might be confused with other characters. If you think that this is intentional, you can safely ignore this warning. Use the Escape button to reveal them.
# Movies, stereo photos and splats
Frame Control has its own Frame-side OpenVR player. It uses SteamOS's ffmpeg
and V4L2 hardware decoder, Python and SteamVR. **No separate player or viewer
is required.** Chromium and immersive WebXR are not in this playback path.
## Use it
In **Tools → Media in the headset**, send a file, choose its layout and press
**Play**. **Theatre** gives it a larger screen and an 85% black surround.
**Stop** removes both. Refresh reads the library and the player's state.
The screen follows your head; it isn't a saved world-space panel.
The command-line route uses the same player:
```sh
scripts/push-vr-video.sh --launch --theatre ~/Movies/film_SBS.mp4
scripts/push-vr-video.sh --launch --layout ou ~/Pictures/stereo.png
scripts/push-vr-video.sh --launch capture.splat
scripts/push-vr-video.sh --list
scripts/push-vr-video.sh --stop
```
Uploads live in `~/Videos/FrameControl/<id>/` on the Frame. Each upload gets
its own directory, so sending another file with the same name doesn't replace
it. The UI's existing upload limit is 8 GiB. Transfers use the app's rsync/scp
path; resumable large uploads are not implemented here yet. Existing
`~/Videos/VR` files and Proton prefixes are left alone. The script no longer
launches DeoVR, links a Proton prefix, or accepts directories.
| Media | Supported preview |
|---|---|
| Movies | H.264 / H.265, mono or left-first SBS / top-first OU; audio through the Frame's PulseAudio-compatible server |
| Stereo photos | PNG / JPEG containing both eyes, SBS or OU |
| Gaussian splats | Common 32-byte `.splat` records, 1–20,000 Gaussians; a stationary stereo preview |
Auto layout reads delimited filename tags: `_SBS`, `_HSBS`, `_LR`, `_OU`,
`_TB`, `_HOU`, `_HTB`, `_FSBS`, `_FOU`, `_FTB`. SBS/OU without `F` means
half-resolution packing. Full packing preserves each eye's original aspect.
It also accepts FFmpeg's `stereo_mode` metadata (`left_right`, `top_bottom`,
`mono`); left/right and top/bottom metadata are treated as full packing.
Choose an explicit layout if that assumption doesn't match the file.
Unknown or conflicting tags ask for a choice, rather than silently flattening
stereo. The explicit selector always wins. Layout is not guessed from resolution.
## What was verified
**Verified remotely, 2026-09-28:** SteamOS 0.4.1, BUILD_ID
`20260925.6191901`, SteamVR 2.18.1. Nobody wore the headset for these checks.
All test media were generated by us. No paid content or DRM was involved.
- Stock `ffmpeg` `h264_v4l2m2m` decoded 150 frames of our 1920×1080 H.264
test in 0.128 s (decode only); converting all frames to RGBA took 0.242 s.
- Our ffmpeg → Python → OpenVR path submitted all 150 frames and exited 0.
The 1280×720 prototype took 4.775 s; the full 1920×1080 player took
4.618 s. These are wall times, not in-headset frame-rate measurements.
Finishing a 5 s clip early showed those probes weren't paced, so the final
player paces output at 30 fps: all 150 frames then completed in **5.025 s**.
- The H.265 hardware decoder also completed the generated 1080p clip
(60 frames, exit 0); this is a short compatibility check, not a 4K/8K benchmark.
- Our actual player, launched through Frame Control's media API, displayed
SBS and OU photos with red only in the left-eye capture and cyan only in
the right. OpenVR's `SideBySide_Parallel` flag separates the eyes; we
rearrange OU rows ourselves.
- A generated 48-Gaussian `.splat` rendered in our own CPU renderer and appeared
in both eyes with separate perspective projections.
- Theatre's owned dark surround and screen appeared in captures. Steam's
dashboard remained available over them. Stop removed the owned overlays
and ended the dedicated user service. No global SteamVR setting was changed.
- A generated H.264/AAC clip reached the Pulse audio output and completed
all 100 video frames with exit 0. This proves the output path, not audible
quality or lip sync.
![Frame Control SBS eye-isolation proof: red left, cyan right](img/media-sbs-proof.png)
![Our Gaussian-splat stereo preview on the Frame](img/media-splat-proof.png)
**Verified failed route:** GStreamer 1.24.2's `playbin` selected
`v4l2h264dec`, delivered the first RGBA sample and then segfaulted (exit 139)
in the basic appsink probe and the OpenVR probe. We do not ship that route.
The ffmpeg path above completed instead.
## Limits and blockers
- **Not verified:** worn-headset comfort, sound quality/lip sync, long movies,
4K/8K decode, HDR, controller interaction or behaviour on other OS builds.
Output is bounded to 1920×1080 packed pixels before OU rearrangement.
- **Not implemented:** pause, seeking, subtitles, playlists, right-eye-first
layouts, non-square pixel correction, VR180/360 projection and fisheye.
This preview is a flat stereo screen, not a dome player.
- **Native spatial-photo blocker:** HEIC/HEIF/AVIF/MPO stereo-container
extraction isn't implemented in our viewer. We reject these rather than
displaying one image and calling it spatial. Export both eyes to PNG/JPEG
first. This is a current implementation gap, not a claim that the Frame
cannot support these containers.
- **Large/immersive splat blocker:** the owned CPU rasterizer projects 3D
covariance into each eye and alpha-composites Gaussians, but renders a
fixed view at 320×240 per eye. It caps each Gaussian footprint at 32 pixels
and normalizes the scene to a two-metre box. Large scenes, PLY/SPZ, live
head-position parallax and navigation need a GPU scene renderer; this
version rejects files above 20,000 records. It is a stereo preview, not
an immersive walk-through.
- A single player can run at a time. It stops at EOF, on **Stop**, or after
four hours in any case, so longer movies are cut off. Still images remain
until Stop or that timeout. There is no delete action yet: remove old
uploads from `~/Videos/FrameControl/` over SSH. The log is
`~/.local/share/frame-control/media/player.log` on the Frame.
- **Panel/stream theatre remains outside this media slice:** SteamVR's
`vrcmd --dock-overlay` accepts `theater`, but docking an existing panel
and its dimming behaviour were not verified here. The shared headset had
workspace and stream tests active. We did not reposition their panels.
Integration with [#22](https://github.com/saphid/frame-control/issues/22)
and [#31](https://github.com/saphid/frame-control/issues/31) can use the
owned rendering hook below; no second stream/window manager is introduced.
## Integration hook
`POST /api/upload` with `X-Mode: media` and `X-Filename` sends a file and
returns its `id`. `POST /api/media` accepts:
```json
{"action":"play","id":"<32 hex characters>/film_SBS.mp4","layout":"auto","theatre":true}
```
Other actions are `list`, `status`, `stop`. These use the normal `X-Frame-UI`
guard. Play reports **starting**, not a claim that frames reached the headset;
read status for `playing`, `ended`, `stopped` or `error`.
The own-code files are copied to `~/.local/share/frame-control/media/` and
run in the `frame-control-media.service` systemd user unit. Stop affects only
that unit, including its decoder child.
For a Frame-side stream producer, `frame_media_player.Overlay` exposes
`create(key, width, distance, stereo=False, aspect=1, order=1)`,
`pixels(handle, rgba, width, height)` and `close()`. Use an owned unique key,
pass interleaved RGBA bytes (left/right halves for stereo) and always close in
`finally`. `create` uses a head-relative transform; it does not move another
app's panel. The player demonstrates a separate black surround overlay.
`frame_media.stereo_pixels` converts OU to SBS. This is the narrow rendering
hook for stream/workspace work; it doesn't capture or manage a Mac/PC stream.
## Optional separate app
You can independently install **DeoVR Video Player** (free Steam app 837380)
if you prefer its VR180/360 features. Earlier tests on SteamOS 0.3.0,
build `20260922.6101926` (2026-09-25), verified its OpenVR initialization and
8K H.265 streamed VR180 decoding under Proton. Frame Control's media features
neither install nor launch it, and do not depend on it. Its behaviour and
file-naming conventions are not evidence about our player.