LED box / CAVE rigs (one eye, one off-axis camera per screen)

Verified Oct 2026: a 5-screen LED room (front, left, right, ceiling, floor) rebuilt in UE 5.8.4 from the C4D rig that rendered the room's precomps.

Find the real rig

• The playback project (e.g. an Unreal VR walkthrough of the room) usually only plays the videos (FileMediaSource → MediaPlayer → MediaTexture). Read the media sources' file paths (strings inside the .uasset) → the delivered videos → their resolutions (tkhd box of the .mp4) — and look next to the render logs for the scene that made them (C4D batch-render XML logs name the take, camera, render setting, resolution, fps).
• An unfolded test video (e.g. 8000x7020 = side+front+side wide, ceiling+front+floor tall) confirms the layout.

Read a C4D rig headless (c4dpy, read-only)

c4dpy g_licenseModel=LICENSEMODEL::MAXONAPP dump.py scene.c4d out.json < /dev/null (see the user's c4dpy memory). Load with SCENEFILTER_OBJECTS; for each CAM* object: parents, GetMl(), GetMg() at a few frames, CAMERA_FOCUS, CAMERAOBJECT_APERTURE (sensor width), CAMERAOBJECT_FILM_OFFSET_X/Y; render data (RDATA_XRES/YRES/FILMASPECT/FRAMERATE); takes (GetCamera(td), GetRenderData(td)). A 750 MB scene loaded in ~40 s.

Film offsets C4D → Unreal (derived from a consistent room, then from UE source)

• C4D: X = fraction of the film WIDTH, + shifts the view right; Y = fraction of the film HEIGHT,
  • shifts the view DOWN. (Checked by rebuilding the room: with this convention every wall edge meets its neighbour within 0.1 %, and each wall's aspect equals its render aspect.)
• UE CineCamera filmback SensorHorizontalOffset / SensorVerticalOffset (mm, 5.8): becomes OffCenterProjectionOffset = 2·offset / sensor size and the projection uses M[2][0] = -offX →
  • looks right / up.
• So h_mm = X · sensor_w, v_mm = −Y · sensor_h, sensor_h = sensor_w · res_y / res_x.
• Orientation C4D (X right, Y up, Z forward) → UE (X forward, Y right, Z up): front (0,0,0), left yaw −90, right yaw +90, ceiling pitch +90 (image up = backwards), floor pitch −90 (image up = forward).

Rig on a rail

• All cameras at one eye point on the mount, relative location 0. Keep the box LEVEL: world rotation = compose_rotators(screen_rotation, Rotator(yaw=rail_yaw)), set with set_actor_rotation after attaching (a rail locked to a climbing spline would tilt the room's horizon). Re-level after moving the rail.
• Eye height matters: in the room the eye sits at the wall-height ratio (e.g. 0.425 × front distance above the floor ≈ 1.8 m); a virtual eye 30 cm above the ground reads as crawling.

One video per screen

• One shot sequence per screen: the main sequence as a SubTrack section + a CameraCut to that screen's camera. The main sequence must have no camera cut — MRQ used the inner sequence's cut over the shot's (all five renders came out as the front view, letterboxed).
• One MoviePipelinePrimaryConfig per screen (AssetTools.create_asset(..., MoviePipelinePrimaryConfig, MoviePipelinePrimaryConfigFactory())) with that screen's output_resolution — one job per screen (a job has one resolution; "render all cameras" can't mix sizes).
• Checking: render one frame per screen (custom playback range, quarter scale) and paste them into the unfolded layout. AutomationLibrary.take_high_res_screenshot(camera=...) ignored the camera while the viewport piloted another one and forced the viewport aspect — not usable for this.
• Quick look through one off-axis camera without rendering: CaptureViewport (90° horizontal FOV, viewport aspect) from the camera's location, pitched to the frame's centre angle (atan(sensor_v_offset / focal)), then crop columns ±(sensor_w/2)/focal and rows to tan range [(-h/2 + v_off)/focal, (h/2 + v_off)/focal] (px = W/2 per tan unit, rows through the pitch). Matched an MRQ frame of the same shot within ~8 %. No DoF, and editor-preview lighting.