Proposal: DisplayXR for Adobe Substance 3D Modeler (closed legacy OpenXR app, dedicated 3D window, Leia SR tested)

[joehillthunder]

Summary

Substance 3D Modeler 1.22.8 runs its VR mode on DisplayXR today (Windows, Vulkan XR_KHR_vulkan_enable v1, one side-by-side swapchain). The status light reads "3D active". Modeler is closed, with no render hook or scripting for the viewport, so it can't do texture mode. Its target class is the dedicated 3D window from the architecture direction, driven by the shared host library through a thin adapter (launcher + rig + coupling list).

The overlay PRs used to get here (#2332 embed + UI peek, #2350 live depth) are test scaffolding only, not a request to merge. #2333 (gamepad Touch emulation), #2334 (rig pitch/pull-back) and #2347 (mouse hand + stylus) carry the input and framing ideas below.

What was measured (the tested display: a 3840×2400 tracked panel, 34.4 cm wide)

• Depth: sim-display atlas capture, clay silhouette disparity.
  • The default legacy rig (ipdFactor 1) put the clay 1.30–1.41 % of view width in front of the glass, about 10× the 0.14 % front comfort budget.
  • Disparity is linear in strength (50 % → 0.73 %, 10 % → 0.16 %).
  • With ipdFactor 0.1 in the per-title rig, the clay sits at 0.16 % front, with nothing behind the glass.
• Framing: the clay starts below a level 1.6 m camera, so the rig needs pitchDegrees −30 and pullbackMeters 0.6 (Proposal: legacy camera profile pitchDegrees + pullbackMeters (starting framing) #2334). With the narrow comfort vFOV (0.355 rad) only the backdrop was visible; 1.05 rad frames the block.
• Input:
  • Gamepad: an Xbox pad emulates two Touch controllers.
  • Mouse: the mouse is a tracked "sculpting hand" under the cursor (sculpt, grab, wheel = depth), plus an MX Ink stylus profile. The hand has to be placed from the views exactly as the app receives them after the rig eye override; the device's head-relative views were 8.7 cm off.
  • Modeler's own desktop-viewport mouse does nothing in VR mode (posted drags vs a no-input control: no change). For a closed VR app, the mouse has to become a tracked device.
• Concurrency: Modeler wove alongside two other hosted apps on one panel at once.

What a closed legacy app needs from the shared host library

1. A dedicated 3D window with dock/undock beside the app, F11 full panel, and a "3D on a second monitor, UI on the 2D one" mode. With the 3D drawn over the app instead, menus and dialogs end up underneath it; the peek workaround is fragile (the 3D comes back after 4 s idle even with a dialog open).
2. Rig seeding from display info, not constants. The nominal viewing distance, comfort budget (Display info: publish the comfort budget (front/behind disparity limits), e.g. layout.comfort {front_frac, behind_frac} #2388) and panel size should seed convergence, vFOV and ipdFactor. A fitter should then scale strength to the measured content disparity. Modeler's 0.1 is a hand-fitted constant for one canvas size and one scene, and it drifts on resize and as the clay moves toward the camera (Focus on camera rigs: convergence isn't a pure glass shift, and the comfort clamp assumes content at infinity #2385).
3. Input adapters as library features: mouse → tracked hand from the final app views, gamepad → Touch, stylus profile, and per-app mapping tables. A closed app can't remap its own controls.
4. A live depth contract with "relative to the title's nominal" semantics, so one shared control works across apps (proved with the DXR_HostedDepthRel scaffolding).
5. Per-instance capture for automated checks when two copies of an app run (Proposal: per-instance atlas capture trigger (displayxr_atlas_trigger.<exe>.<pid>) — two copies of one app collide under windowed weaving #2351).

Constants that are named gaps

Glass 0.8 m, vFOV 1.05 rad, pitch −30°, pull-back 0.6 m, ipdFactor 0.1, menu inset 33.4 DIP. Each should come from display info, or from the app's own window and camera, through the host library.

Code

A private adapter repository: launcher, rig JSON, a coupling module listing every Modeler/runtime internal relied on (each marked "not a contract"), and a status-label script. Apache-2.0. Available on request.

🤖 Generated with Claude Code

[joehillthunder]

Final update (2026-10-27)

#2393 is fixed in v2.55.0. Modeler was re-tested without DXR_WEAVE_SINGLE_THREAD: Vulkan XR_KHR_vulkan_enable v1, with the VK_LAYER_DXR_queue_lock repaint.

• Released v2.55.0: 9 minutes. Dev v2.55.0-33, embedded in the viewer: 6.5 minutes.
• No app-frame acquire timeouts. The runtime logs single weave-thread ownership as auto-enabled, and the env var is gone from the launcher.
• One resize to 5040×2900 gave VK_ERROR_OUT_OF_DATE_KHR on present; the swapchain was recreated (VK: DXR_WEAVE_REPAINT_FORCE=1 plus repeated fullscreen transitions loses the device (still-in-use on one cmd buffer + one fence) #911).

Released runtimes reject unknown camera-rig keys. v2.55's u_camera_profile.c drops the whole profile when it sees the pitchDegrees / pullbackMeters keys from #2334.

• The launcher now picks the rig per runtime, with a -CameraRig <json> override.
• The released rig is level and can't frame Modeler's clay, because the camera starts inside it. Framing needs Proposal: legacy camera profile pitchDegrees + pullbackMeters (starting framing) #2334.

Depth strength (ipdFactor), measured by atlas capture. The comfort budget is 0.14 % in front and 0.42 % behind.

• 0.1 (the clay comfort fit): 0.16 % in front, but imported content reads flat. A 60 cm cloth gave 0–11 px of 1920.
• 0.3, on CC0 marble bust + chess set (~55 cm), dev rig: everything sits behind the glass, median 0.31 %, p95 0.47 % of view width. 0.25 fits the budget.

One constant can't serve both the clay and imported scenes, which supports a content-bounds + display-info fitter (#2385). Fill is still about 20 % of the view.

Gaps found on the panel:

• Esc closes the session. Pressing Esc in the runtime's 3D window (non-workspace mode) closes the session, so Modeler leaves 3D. That conflicts with apps that use Esc themselves.
• No automation route into a closed app. A posted WM_DROPFILES and a mesh path on the command line are both ignored while 3D is on, so content import is a manual step.

Repo: joehillthunder/displayxr-substance-modeler (private, latest 17e50b6). It has launcher hygiene, the per-runtime rig, -CameraRig, gamepad-dev setup (allowlist + per-app pin), Enter-resets-view and the softened slider. The target class is unchanged: a dedicated 3D window plus the host library, once its design lands.

🤖 Generated with Claude Code

[joehillthunder]

Done (2026-10-28)

• Readiness on the released runtime: installed v2.55.0 with the Leia SR plug-in 2.18.3. The log shows the Program Files runtime and active plug-in: id=leia-sr, and Modeler exits cleanly.
• Framing: the released rig can't frame Modeler yet. The camera starts inside the clay because Proposal: legacy camera profile pitchDegrees + pullbackMeters (starting framing) #2334 isn't released (still missing in v2.57.1), so on the panel we use the dev runtime for now.
• Depth (corrected): the earlier dev rig (pull-back 0.4, strength 0.1) was Over budget: it measured −0.31 to −0.42 % in front across three later launches, 2.2–3× the 0.14 % front budget; Modeler places the clay relative to the viewer at start, so depth varies per launch. New default rig (pull-back 0.5, ipd 0.06; dev runtime with the installed Leia SR plug-in 2.19.1): median −0.16 % in front, 1.1× the 0.14 % front budget, ~30 % fill, weave 60 Hz. Seated check: Pass at 1.1× the front budget.
• Imports: auto-fit keeps imported meshes framed for the 3D screen (now pull-back only). A large rock set fits at a middle size (+0.42 %), and holding S in the 3D view pulls back further.
• USD: Modeler's own USD of an imported set opens in USD Composer textured (6 meshes, UsdPreviewSurface + OpenPBR) and renders in RTX.
• Input: a gamepad is no longer a requirement. Emulated controller hands have fixed poses, so a pad can't aim a sculpt; the mouse hand is the input to use.