AutoGineer
Generated banner for Omniverse Realtime Viewer
Generated logo for Omniverse Realtime Viewer

Omniverse Realtime Viewer

NVIDIAOfficialCreated by NVIDIA Omniverse

  • Skill
  • Tool integration
  • Software development
  • Apache-2.0
  • Medium risk
Download ZIP

From the author

Use as the top-level router for Omniverse Realtime Viewer USD apps, XR delivery, SPG GPU processing, stylized rendering, and focused viewer reference documents.
— the description by NVIDIA Omniverse

AutoGineer summary

Written by AutoGineer from the skill’s own files, not by the author

The Omniverse Realtime Viewer skill acts as a central routing and architectural framework for developing USD-based 3D viewer applications. It provides a structured approach to composing rendering, scene state, and delivery mechanisms, ensuring that developers can integrate OVRTX rendering, OVStage scene management, and various UI or streaming adapters effectively. The skill emphasizes a modular architecture where the viewer runtime, UI/input adapters, and delivery systems are decoupled, allowing for consistent behavior across local and streamed environments.

What it does

  • Provides a routing architecture for USD viewer apps, XR delivery, and GPU-accelerated rendering.
  • Defines the separation of concerns between viewer runtime, UI adapters, and delivery systems.
  • Offers guidance on integrating OVRTX for high-fidelity 3D rendering.
  • Establishes conventions for camera interaction, scene state management, and input handling.
  • Supports the implementation of specialized features like stylized rendering and CAE/CFD visualization.
  • Provides protocols for agent-to-UI communication and scene control.

When to use it

  • When starting the development of a new USD-based realtime viewer application.
  • When integrating advanced rendering features like stylized visualization or AOV switching.
  • When building XR or spatial editing capabilities into an existing Omniverse application.
  • When you need to standardize the architecture for local and streamed viewer deployments.
  • When implementing chat-driven scene control or agent-based inspection of a 3D scene.

Advantages

  • Promotes a clean, modular architecture that separates rendering logic from UI and transport.
  • Ensures consistent behavior across different delivery methods, such as local windows or WebRTC streaming.
  • Provides clear paths for extending functionality without disrupting the core viewer runtime.
  • Facilitates the integration of complex simulation and visualization data into the viewer.
  • Standardizes conventions for scene state, selection, and camera control.

Usage examples

“How do I set up a basic USD viewer application using the Omniverse Realtime Viewer architecture?”
The agent will guide you to start with the viewer app reference, resolve dependencies, and implement the core runtime owner to manage the OVStage and OVRTX renderer.
“I need to add stylized rendering to my existing viewer. Where should I start?”
The agent will direct you to the stylized visualization references to implement the required USD/Lua/CUDA nodes within your existing runtime owner.
“How can I enable an AI agent to control the viewer scene?”
The agent will point to the A2UI Scene Copilot reference to implement an operation interface that allows the agent to send commands through the existing runtime owner.
“How do I switch between different render outputs like beauty and depth in my viewer?”
The agent will explain how to use the AOV switching reference to map RenderVars and update the RenderProduct's orderedVars through the runtime owner.

Install

  1. Download the skill as a ZIP:

    https://autogineer.com/agent-skills/nvidia/omniverse-realtime-viewer/download

  2. Unzip it into the cross-client skills folder of your project:

    mkdir -p .agents/skills && unzip omniverse-realtime-viewer.zip -d .agents/skills

Documentation by NVIDIA Omniverse

<!-- SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. -->

<!-- SPDX-License-Identifier: Apache-2.0 -->

Omniverse Realtime Viewer

Build responsive USD viewers by composing OVRTX rendering, OVStage scene state, and the delivery/UI libraries appropriate to the app. Use this package for viewer architecture and interaction; use each selected SDK's upstream guidance for installation and exact release-specific APIs.

Start With The App

  1. When delivery is explicit, select its recipe through references/routing.md and follow its current phase. Otherwise start with references/usd-viewer-app/README.md. For an existing app, keep its architecture and read only the references for the requested change. Broad app, conventions and UI catalogs are lookups as needed.

  2. Resolve the selected libraries through references/dependencies/README.md. Preserve the project's manifest unless an upgrade is requested.

  3. Build a working frame and camera interaction in the chosen shell, then add the requested capabilities. Read references/conventions.md for shared input, camera, selection, and protocol conventions.

  4. Use references/viewer-ux-workflow/README.md for broader interface design. Loading, selection feedback, errors, and input responsiveness belong in the first working app, alongside rendering.

  5. Validate the actual delivery path with references/validation.md. A browser screenshot does not validate a native window or a different transport.

Architecture

OwnerResponsibility
Viewer runtimeOwn the live OVStage, publication clock, OVRTX renderer, scene changes, and render stepping.
UI and input adapterDisplay frames and copied app state; send user intent to the runtime owner.
Storage/query/simulation workersPerform their assigned work and return results through the selected integration boundary.
Delivery adapterPresent locally or stream through ovstream; respect frame ownership and backpressure.

Use OVRTX for USD/3D rendering. Browser apps display its WebRTC video; local apps present its output through their native shell. A missing GPU means rendering is unverified, not a reason to substitute a browser 3D renderer.

Keep viewer cameras, settings, effects, and transient edits in app-owned session or runtime state so the source USD stays intact. Serialize scene mutation and render stepping through the runtime owner. Read references/ovstage-ovrtx-integration/README.md for that loop, then the population or data-plane references when implementing scene loading or attribute writes.

Keep expensive loading and metadata work off the UI and render threads. Keep transport details behind a delivery adapter so camera, selection, properties, and loading state can be reused across local and streamed shells.

Add Capabilities As Needed

Use references/routing.md to select camera/picking, scene management, render settings, cloud assets, simulation, editor tools, deployment, or CAE/CFD visualization. Read the selected capability's implementation flow and its linked contract; the full reference catalog is not a prerequisite for a first frame.

For simulation, use the routing reference's Simulation Ecosystem section plus references/realtime-sim-integration/README.md. For UI-driven forces or control, add references/interaction-round-trip/README.md.

For example scenes, use references/usd-sample-data/README.md. When a runtime failure needs diagnosis, use references/troubleshooting/README.md and the selected SDK's matching release guidance before changing the architecture.

Solid Modeling: Upstream First

For usd-brep, exact BRep, solid modeling, kernel failures, CAD topology, or BrepArray questions, first read references/dependencies/README.md → USD BRep. Follow the owning repository's current agent guidance, skills, operation docs, and package API before selecting local integration references. Discover the upstream router rather than assuming skill names or directory layouts remain fixed. Upstream owns modeling semantics, schema rules, tuning, and diagnostics; this package owns OVRTX display, worker handoff, selection mapping, and publication timing.

GPU Processing And Stylized Viewers

For custom GPU processing, chained outputs and lens sampling, read references/spg-processing/README.md. For toon, ink, technical/X-Ray styles, geometry strokes or full wireframe, continue to references/stylized-visualization/implementation.md. Its USD/Lua/CUDA starting node and geometry/depth/composite build sequence work with the selected viewer recipe; use references/stylized-visualization/style-recipes.md for the preset math and references/stylized-visualization/README.md for supporting contracts. Implement the result inside the existing OVRTX/runtime owner and deliver the published processed output. These instructions do not require a sample checkout or a prior stylized-rendering implementation.

For an optional usd-brep geometry source, first read references/dependencies/README.md → USD BRep and the owning repository's current agent guidance, skills, and operation docs; then read references/stylized-visualization/usd-brep-source.md for the display boundary.

XR Viewers

For headsets, immersive viewers, or spatial editors, start with references/xr-viewer-recipe/README.md. Resolve SDK setup and contracts through references/dependencies/ovxr.md; the library owns OpenXR lifecycle and frame submission. For controllers, hands, navigation, or scene tools, add references/xr-spatial-input/README.md and references/ovxr-python-interactions/README.md. For projection-integrated UI, read references/ovxr-composition-ui/README.md and references/ovxr-composition-ui-validation/README.md.

Files

references/cpp-native-viewer/presentation-loop.md

C++ Presentation And Render Loop

Owned Frame Transfer

The runtime owner CPU-maps LdrColor, copies its pixels, unmaps, and destroys the step result before publishing the owned frame. It never calls OpenGL. Respect the DLPack byte offset and strides; mapped pointers never cross the thread boundary.

#include <cstdint>
#include <vector>

struct DisplayFrame {
    int width = 0;
    int height = 0;
    std::vector<std::uint8_t> rgba;
};

static bool copyRgba(DisplayFrame& frame, const DLTensor& tensor)
{
    if (!tensor.data || tensor.device.device_type != kDLCPU ||
        tensor.ndim != 3 || !tensor.shape || tensor.dtype.code != kDLUInt ||
        tensor.dtype.bits != 8 || tensor.dtype.lanes != 1 ||
        tensor.shape[0] <= 0 || tensor.shape[1] <= 0 || tensor.shape[2] != 4)
        return false;
    const int height = static_cast<int>(tensor.shape[0]);
    const int width = static_cast<int>(tensor.shape[1]);
    const int64_t row = tensor.strides ? tensor.strides[0] : width * 4;
    const int64_t pixel = tensor.strides ? tensor.strides[1] : 4;
    const int64_t channel = tensor.strides ? tensor.strides[2] : 1;
    if (row <= 0 || pixel <= 0 || channel <= 0) return false;
    const auto* source = static_cast<const std::uint8_t*>(tensor.data) + tensor.byte_offset;
    frame.width = width;
    frame.height = height;
    frame.rgba.resize(static_cast<size_t>(width) * height * 4);
    for (int y = 0; y < height; ++y)
        for (int x = 0; x < width; ++x)
            for (int c = 0; c < 4; ++c)
                frame.rgba[(static_cast<size_t>(y) * width + x) * 4 + c] =
                    source[y * row + x * pixel + c * channel];
    return true;
}

Publish only a successfully copied complete frame, with its stage generation and dimensions, into a single latest-frame slot. The UI takes shared ownership or moves that buffer under a short lock, then releases the lock before upload.

Native Frame Order

For each changed scene state, write through OVStage at ordinal N, wait for writes, and advance/wait the write floor to N. Before rendering, call and wait ovrtx_update_from_stage(renderer, N). The C function ovrtx_step_with_stage() requires this update; unlike Python Renderer.step(), it does not supply the high-level wrapper's automatic update.

After the update, apply dirty native outline/pickability state, then enqueue a pending pick. Population can recreate renderer-owned prim state, so retained selection must be dirty after a same-scene replacement and applied here, once the new population is visible. A pick decoded after this frame updates selection intent for the next frame.

The helper below assumes the update has completed and optional renderer-native commands have been enqueued. It performs one step and maps the display output; decode OVRTX_RENDER_VAR_PICK_HIT in this same result iteration when a query was queued (see interaction-features.md).

static bool sameString(ovx_string_t value, const char* literal)
{
    const std::string expected(literal);
    return value.ptr &&
        value.length == expected.size() &&
        std::char_traits<char>::compare(value.ptr, expected.data(), expected.size()) == 0;
}

static const DLTensor* firstTensor(const ovrtx_render_var_output_t& output)
{
    if (!output.tensors || output.num_tensors == 0) return nullptr;
    return output.tensors[0].dl;
}

static bool stepRenderAndCopy(
    ovrtx_renderer_t* renderer,
    const std::string& renderProductPath,
    double deltaSeconds,
    ovstage_ordinal_t ordinal,
    DisplayFrame& frameCopy)
{
    frameCopy = {};
    bool copied = false;
    const ovx_string_t productPath = toOvxString(renderProductPath);
    const ovrtx_render_product_set_t products = {&productPath, 1};

    ovrtx_step_result_handle_t stepHandle = OVRTX_INVALID_HANDLE;
    const ovrtx_enqueue_result_t step =
        ovrtx_step_with_stage(renderer, products, deltaSeconds, ordinal, &stepHandle);
    if (!ok(step) || stepHandle == OVRTX_INVALID_HANDLE) {
        printLastOvrtxError("Failed to step OVRTX");
        return false;
    }

    ovrtx_render_product_set_outputs_t outputs = {};
    const ovrtx_result_t results =
        ovrtx_fetch_results(renderer, stepHandle, ovrtx_timeout_infinite, &outputs);
    if (!ok(results) || outputs.status != OVRTX_EVENT_COMPLETED) {
        printLastOvrtxError("Failed to fetch OVRTX results");
        ovrtx_destroy_results(renderer, stepHandle);
        return false;
    }

    for (size_t productIndex = 0; productIndex < outputs.output_count; ++productIndex) {
        const auto& product = outputs.outputs[productIndex];
        for (size_t frameIndex = 0; frameIndex < product.output_frame_count; ++frameIndex) {
            const auto& frame = product.output_frames[frameIndex];
            for (size_t varIndex = 0; varIndex < frame.render_var_count; ++varIndex) {
                const auto& var = frame.output_render_vars[varIndex];
                // Compare output identity with the full authored RenderVar path.
                // `source_name` classifies the AOV; it is not the map key.
                if (!sameString(var.render_var_path, (renderProductPath + "/LdrColor").c_str())) continue;

                ovrtx_map_output_description_t mapDesc = {};
                mapDesc.device_type = OVRTX_MAP_DEVICE_TYPE_CPU;
                ovrtx_render_var_output_t mapped = {};
                const ovrtx_result_t map =
                    ovrtx_map_render_var_output(renderer, var.output_handle,
                        &mapDesc, ovrtx_timeout_infinite, &mapped);
                if (ok(map) && mapped.status == OVRTX_EVENT_COMPLETED) {
                    if (const DLTensor* tensor = firstTensor(mapped)) {
                        copied = copyRgba(frameCopy, *tensor);
                    }
                }
                if (mapped.map_handle != OVRTX_INVALID_HANDLE) {
                    ovrtx_unmap_render_var_output(renderer, mapped.map_handle, {});
                }
            }
        }
    }

    ovrtx_destroy_results(renderer, stepHandle);
    return copied;
}

Use scope guards in the app to unmap and destroy results on error/exception paths as well. Publish an error if fetch/map/copy fails instead of reporting a new ready frame. Never step twice just to decode the pick for a displayed frame.

UI Presentation And Input

Use the selected toolkit's event loop and pixel/image presentation API to show the latest complete owned frame independently of SDK progress. Keep toolkit objects on their required thread; copy pixels before releasing the SDK map. The viewport's actual displayed image rectangle owns coordinate conversion, including aspect ratio, letterboxing and display scaling. Toolkit choice does not change native picking or renderer ownership.

For the GLFW/ImGui example, poll GLFW, build ImGui controls, and present the latest frame. Create and resize the GL texture on this thread, set GL_UNPACK_ALIGNMENT to 1, and upload the packed RGBA buffer with glTexSubImage2D(..., GL_RGBA, GL_UNSIGNED_BYTE, frame.rgba.data()). LdrColor is RGBA; do not swap red/blue.

In that example, fit ImGui::Image() inside the available content rect while preserving the rendered aspect ratio. Retain that actual image rectangle for picking and camera input. Reject letterbox clicks; convert accepted image coordinates into the render product's normalized coordinates. Use the same top-left orientation for display and input, and account for GLFW window versus framebuffer scaling.

Route input through the viewport only when the image owns it. Text editing and modal controls consume their events, and gizmo handles get first refusal before camera or picking. Track a pressed drag until release even outside the image; clear drag state when focus is lost. See viewer-input-routing.

Mark readiness only after the first successfully rendered owned frame has been uploaded and swapped while the UI can accept input. During scene preparation continue to present the active scene and loading/error status. An SDK wait remains on the runtime owner; UI polling and normal-close intent remain responsive.

Requirements and compatibility

Works with: Claude Code, Claude apps, OpenAI Codex, GitHub Copilot, Cursor, Gemini CLI, Kiro, OpenCode, Goose, Amp, Roo Code, JetBrains Junie, Factory Droid, Devin (ex-Windsurf), Mistral Vibe, OpenHands, Letta

“Works with” means the tool reads the open Agent Skills format; a skill written for one tool may use features others lack.

Needs: Needs Python, Needs Node.js

The author says: Orchestrator skill. Downstream focused references may require NVIDIA GPUs, ovrtx, ovstream, ovui, OpenUSD, Python, Node/React, Tauri, Electron, C++, or cloud GPU deployment access depending on the selected viewer path.

Security

What our scanner noticed· Medium risk

  • Large base64-encoded blobs
  • Instructions that use the network

A static scan is not a guarantee. Skills are third-party content: review a skill before you let an agent use it.

Attribution and licence

Created by
NVIDIA Omniverse
Publisher
NVIDIA (Official)
Source
NVIDIA/skills · skills/omniverse-realtime-viewer · commit fce3cca
Licence
Apache License 2.0 (redistribution allowed)
  • LICENSE (added by AutoGineer from LICENSE-APACHE)
  • ATTRIBUTION.md (written by AutoGineer)
Synced

AutoGineer is not affiliated with or endorsed by NVIDIA Omniverse or NVIDIA.

Download ZIP

The ZIP holds the skill’s files, its licence files and an ATTRIBUTION.md. Download omniverse-realtime-viewer.zip