Outputs

The tables below list the RenderVar source names available for camera sensors, grouped by render mode. Shapes are channel-last DLPack tensor shapes.

Use the sourceName value as the sourceName attribute on a RenderVar prim. Click any example image to enlarge it. Buffer examples are decoded illustrations, not raw image outputs.

Bounding boxes, occlusion, identifier maps, instance maps, and stable-to-semantic ID pairs expose named, typed tensors through ovrtx_render_var_output_t. Field tensors may be strided views of one allocation. Semantic boxes also expose valid_count on the requested device; only the first valid_count[0] rows are populated. Keep the mapping alive while using its tensors. Token-map range and tile metadata are CPU parameters; CameraParams contains only CPU parameters. Structured outputs use schema version 0.

Note

The example images for non-color outputs have been remapped for easier visualization and do not represent raw output values. All output example images in the table were generated from the semantic-segmentation example.

Real-Time Path Tracing and Minimal use 50 warmup frames; Path Tracing uses one frame at 128 samples per pixel after an initial step that creates instances for selection. The robots then move for a separate motion capture, whose hue shows direction and brightness shows displacement (20 pixels reaches full brightness).

Bounding-box overlays show only robot geometry: raw box outputs show individual parts, while semantic box outputs show one aggregate box per top-level robot. Colors distinguish robots. 3D corners are transformed to world space and projected using CameraParams.

SelectionState previews decode each of the four coverage samples before averaging their colors. Selected robots use cyan (group 1), orange (group 2), and purple (group 3); visible samples have full brightness, hidden samples 25%, and background is black. These colors illustrate data, not renderer styling.

Occlusion colors loose part boxes from green (visible) to red (occluded); invalid ratios are omitted.

ID-map examples pair an annotated scene with decoded metadata; delta examples use their first-frame entries.

Real-Time Path Tracing

sourceName	Format	Type	Shape	Example
LdrColor	RGBA	uint8	(H, W, 4)
Tone-mapped color in sRGB space. Standard final image output suitable for display or saving as PNG/JPEG.
HdrColor	RGBA	float16	(H, W, 4)
Linear-space HDR color preserving full scene luminance (the beauty output; see the Glossary). Use for post-processing, compositing, or linear-space workflows. Delivered as float16 unless the RenderProduct sets omni:rtx:rt:fp32Target, which delivers float32; see HdrColor precision.
SelectionState	Four coverage samples	uint16	(H, W, 4)
Selected-surface classification at output resolution. Each channel stores group in bits 0–7 (zero means background) and visibility in bit 8 (1 visible, 0 hidden); bits 9–15 are zero. Channels are adaptive coverage samples, not RGBA colors. Only the nearest selected surface is represented; objects sharing a group form one mask. Requesting this output replaces the built-in selection overlay for this RenderProduct. XR foveation unwarp is unsupported: the output is cleared and a warning is logged. Disable unwarp to read selection in warped coordinates. See the selection SPG example for custom outlines, fills, and hidden-surface cross-hatching.
NormalSD	XYZA	float32	(H, W, 4)
World-space surface normals. Python/C decoding.
DepthSD	Z	float32	(H, W, 1)
Deprecated and still delivered; use DistanceToImagePlaneSD or DistanceToCameraSD instead. For a perspective camera, unitless raster depth mapped from 1 (near clip plane) to 0 (far clip plane); an orthographic camera reverses the mapping, and a fisheye or other lens projection stores the distance from the camera. A pixel whose ray hit nothing holds the far-clip value, which is 0 for a perspective camera. Python/C decoding.
DistanceToCameraSD	Z	float32	(H, W, 1)
Euclidean distance from the camera origin to each surface point, in meters whatever the stage’s metersPerUnit. A pixel whose ray hit nothing holds infinity. Python/C decoding.
DistanceToImagePlaneSD	Z	float32	(H, W, 1)
Perpendicular distance from the image plane to each surface point, in meters whatever the stage’s metersPerUnit. A pixel whose ray hit nothing holds infinity. Python/C decoding.
DiffuseAlbedoSD	RGBA	uint8	(H, W, 4)
Diffuse surface albedo (base color without lighting). Python/C decoding.
Camera3dPositionSD	XYZA	float32	(H, W, 4)
Camera-space 3D position of each visible surface point, in scene units, with W = 1. A pixel whose ray hit nothing has W = 0 and holds a point far along the ray, so test W rather than the position magnitude to find background; a min/max-stretched preview of the raw values comes out flat. Python/C decoding.
TargetMotionSD	XYZW	float32	(H, W, 4)
Backward screen-space displacement in pixels: add XY to the current pixel position to find its previous-frame position. Z is the previous-frame occlusion flag; W indicates a geometric hit within the occlusion-check range. Invalid correspondences are (-1, -1, -1, -1). Python/C decoding.
SemanticSegmentation	ID	uint32	(H, W, 1)
Per-pixel semantic ID. Decode SemanticIdMap to map each integer ID to its semantic label string. Class index 1 marks unlabeled geometry. Class indices occupy the first word of the four-word class ID in SemanticIdMap, so the matching class tuple is (1, 0, 0, 0). Python/C decoding.
SemanticIdMap	IdentifierMap	Named tensors	id (N, 4), label_offsets (N+1,), label_data (B,)
Semantic IDs and UTF-8 labels as uint32 IDs, uint64 offsets and uint8 label bytes. Row order, sparse empty rows and all four ID words are preserved. Python/C decoding.
SemanticInstanceSegmentation	ID[4]	uint32	(H, W, 4)
Stable ID of the semantic prim associated with each pixel. Maintain a prim-path lookup from StableIdMapDeltas. Background is all zeros; unlabeled hits are (0, 0, 0, 1). Python/C decoding.
StableIdSegmentation	ID[4]	uint32	(H, W, 4)
Stable geometry-instance ID per pixel. The four words form one identifier, including its subset component. Background is all zeros. Python/C decoding.
NonStableInstanceSegmentation	ID	uint32	(H, W, 1)
Runtime semantic-instance index plus 2 per pixel. Zero is background; 1 is an unlabeled hit. These indices can change as scene contents change. Python/C decoding.
BoundingBox2DTightSD	BoundingBox2D	uint32, int32	Named tensors
Per-instance visible-pixel bounds: instance_id, render_tile_id and extent (N,2,2). Reject inverted bounds. Python/C decoding.
BoundingBox2DLooseSD	BoundingBox2D	uint32, int32	Named tensors
Per-instance bounds including occluded geometry: instance_id, render_tile_id and extent (N,2,2). Python/C decoding.
BoundingBox3DSD	BoundingBox3D	uint32, float32	Named tensors
Per-instance local extent (N,2,3), object_to_world (N,4,4), instance_id and render_tile_id. Python/C decoding.
SemanticBoundingBox2DTight	SemanticBoundingBox2D	uint32, int32	Named tensors
Visible-pixel bounds grouped by semantic instance: id, semantic_id and extent (N,2,2). Consume only valid_count[0] rows. Python/C decoding.
SemanticBoundingBox2DLoose	SemanticBoundingBox2D	uint32, int32	Named tensors
Bounds including occluded geometry grouped by semantic instance: id, semantic_id and extent (N,2,2). Consume only valid_count[0] rows. Python/C decoding.
SemanticBoundingBox3D	SemanticBoundingBox3D	uint32, float32	Named tensors
Local extent (N,2,3), object_to_world (N,4,4), id and semantic_id. Consume only valid_count[0] rows. Python/C decoding.
OcclusionSD	Occlusion	uint32, float32	Named tensors
instance_id, render_tile_id, and occlusion_ratio have shape (N,). Ratios range from 0 to 1; -1 means no sampled hits. Python/C decoding.
InstanceIdTokenMapSD	InstanceIdTokenMap	uint32, uint64	Named tensors + CPU params
instance_id, prim_token, prototype_index, and instance_index have shape (N,). CPU scalar params provide min_instance_id, max_instance_id, and render_tile_count. Tensor rows contain no header. Python/C decoding.
InstanceMap	InstanceMap	uint32	semantic_instance_index (N,)
Runtime-instance-to-semantic-instance lookup. 0xFFFFFFFF means no semantic parent. Indices address semantic-instance records, not SemanticIdMap label IDs. Python/C decoding.
StableIdMap	IdentifierMap	Named tensors	id (N, 4), label_offsets (N+1,), label_data (B,)
Stable IDs and USD prim paths as uint32 IDs, uint64 offsets and uint8 UTF-8 bytes. Repeated IDs retain their separate rows. Python/C decoding.
StableIdMapDeltas	IdentifierMap	Named tensors	id (N, 4), label_offsets (N+1,), label_data (B,)
New stable-ID/prim-path entries since the last generated delta, using the named identifier-map tensors. The first frame includes scene-creation entries; an empty delta has zero rows and label_offsets = [0]. Python/C decoding.
SemanticIdMapDeltas	IdentifierMap	Named tensors	id (N, 4), label_offsets (N+1,), label_data (B,)
New semantic-ID/label entries since the last generated delta, using the same named tensors as SemanticIdMap. An empty delta has zero rows and label_offsets = [0]. Python/C decoding.
CameraParams	CameraParams	Named CPU params	No tensors
View and projection matrices, scene-unit scale, resolution, clipping range, camera-model name and lens fields as named CPU parameters, for every requested mapping device. Python/C decoding.
StableIdSemanticIdMap	StableIdSemanticIdMap	uint32	Named tensors
stable_id and semantic_id both have shape (N,4). Preserve all four words and repeated rows when joining semantic instances to their labels. Python/C decoding.

Path Tracing

sourceName	Format	Type	Shape	Example
LdrColor	RGBA	uint8	(H, W, 4)
Tone-mapped sRGB color; uint8 channels decode to [0, 1]. Python/C pixel decoding.
HdrColor	RGBA	float16	(H, W, 4)
Linear-space HDR color preserving full scene luminance (the beauty output; see the Glossary). Use for post-processing, compositing, or linear-space workflows. Delivered as float16 unless the RenderProduct sets omni:rtx:rt:fp32Target, which delivers float32; see HdrColor precision. Python/C pixel decoding.
PtDirectIllumation	RGBA	float16	(H, W, 4)
Direct diffuse reflection. The source name intentionally uses the renderer spelling Illumation. See Compositing and Back to Beauty and Python/C pixel decoding.
PtGlobalIllumination	RGBA	float16	(H, W, 4)
Indirect diffuse reflection. Excludes the glossy reflection and transmission passes. Python/C pixel decoding.
PtReflections	RGBA	float16	(H, W, 4)
Direct and indirect glossy/specular reflection, including direct highlights. Python/C pixel decoding.
PtRefractions	RGBA	float16	(H, W, 4)
Direct and indirect glossy/specular transmission. Python/C pixel decoding.
PtSubsurfaceScattering	RGBA	float16	(H, W, 4)
Direct and indirect diffuse transmission, including the subsurface-scattering contribution. Python/C pixel decoding.
PtSelfIllumination	RGBA	float16	(H, W, 4)
Camera-visible emission, lights, and environment misses. Includes the visible background. Python/C pixel decoding.
PtVolumes	RGBA	float16	(H, W, 4)
Direct and indirect volume-scattering contribution. Include in the seven-pass sum. The captured RGB values are zero in this scene.
PtBackground	RGBA	float16	(H, W, 4)
Separate background output for background/matte workflows. Do not add it to the seven-pass sum: visible background is already included. The captured RGB values are zero in this scene.
PtPreDenoisedResult	RGBA	float16	(H, W, 4)
Accumulated radiance before denoising. Python/C pixel decoding.
PtWorldNormal	RGBA	float32	(H, W, 4)
World-space geometric normal in XYZ, averaged across samples; edges can have length below one. Display [-1, 1] as [0, 1]. Python/C pixel decoding.
PtWorldPos	RGBA	float32	(H, W, 4)
World-space hit position in XYZ, in scene units. The preview rescales each coordinate. Python/C pixel decoding.
PtViewNormal	RGBA	float16	(H, W, 4)
View-space normal in XYZ, averaged across samples. Display [-1, 1] as [0, 1]. Python/C pixel decoding.
PtZDepth	RGBA	float16	(H, W, 4)
Ray-hit distance normalized linearly over the depth range from omni:rtx:pathtracing:zDepthMin to omni:rtx:pathtracing:zDepthMax (render settings on the RenderProduct; 0.1 and 10000 scene units by default) and clamped to [0, 1]; repeated in RGB. Python/C pixel decoding.
PtDiffuseFilter	RGBA	float16	(H, W, 4)
Diffuse material color filter in RGB. Python/C pixel decoding.
PtReflectionFilter	RGBA	float16	(H, W, 4)
Reflection material color filter in RGB. Python/C pixel decoding.
PtRefractionFilter	RGBA	float16	(H, W, 4)
Transmission material color filter in RGB. Python/C pixel decoding.
PtSubsurfaceFilter	RGBA	float16	(H, W, 4)
Subsurface material albedo at the first camera hit, estimated from volume absorption and scattering coefficients and weighted by diffuse transmission. PtSubsurfaceScattering also includes diffuse transmission without volume scattering, so it can be nonzero while this filter is zero, for example with thin-walled materials. Like the other filters, this is not an additive beauty pass. The captured RGB values are zero in this scene.
PtLuminance	RGBA	float32	(H, W, 4)
RGB radiance in RGB rendering; spectral rendering stores CIE Y in each RGB channel. The preview log-compresses values and scales them to the 99th percentile. Python/C pixel decoding.
PtIlluminance	RGBA	float32	(H, W, 4)
RGB irradiance at the first surface, evaluated with a unit Lambertian material. The preview log-compresses values and scales them to the 99th percentile. Python/C pixel decoding.
PtMotion	RGBA	float32	(H, W, 4)
XY screen displacement in pixels; Z change in view-space depth. W is 1. Motion is captured after moving the robots. Python/C pixel decoding.
PtMultiMatte0 – PtMultiMatte7	RGBA	float16	(H, W, 4)
Each RGBA component holds independent coverage for a selected geometry or material ID, from 0 to 1. Configure omni:rtx:pt:multimatte:rId, gId, bId, aId and their rIsMat, gIsMat, bIsMat, aIsMat flags on each RenderVar. Omitted IDs and -1 disable components. Requests determine allocation, up to eight outputs / 32 masks. Geometry IDs use primvars:multimatte_id; material IDs use multimatte_id on the material. The preview combines RGB slices of outputs 0–3; the gray strip shows their full-frame alpha masks. Labels list RGBA selectors (G = geometry, M = material). See Python: Semantic Segmentation and Python/C pixel decoding.
Light Path Expressions	RGBA	float16	(H, W, 4)
Up to nine expression outputs, authored with sourceType = "lpe" and the expression as sourceName. The preview combines vertical slices from slots 0–8, left to right: E.*L, E<RD>.*L, E<RG>.*L, E<TD>.*L, E<TG>.*L, E<RD>L, E<RG>L, E<RD>.+L, E<RG>.+L. See Light Path Expressions and Python/C pixel decoding.
BumpNormal	RGBA	float32	(H, W, 4)
World-space shading normal, including material bump, in XYZ. Python/C pixel decoding.
BumpNormalResearch	RGBA	float16	(H, W, 4)
Half-precision shading-normal diagnostic. Python/C pixel decoding.
DiffuseAlbedo	RGBA	uint8	(H, W, 4)
Diffuse material RGB color, stored as UNORM bytes. Python/C pixel decoding.
SpecularAlbedo	RGBA	uint8	(H, W, 4)
Specular material RGB color, stored as UNORM bytes. Python/C pixel decoding.
DepthLinearized	Scalar	float32	(H, W, 1)
Linear view-space depth in scene units. Ignore the float32 maximum-value miss sentinel. Python/C pixel decoding.
Motion2dXYZ	RGBA	float32	(H, W, 4)
XY screen displacement in pixels and Z view-depth change; W is 1. Python/C pixel decoding.
Motion2dDilated	XY	float32	(H, W, 2)
Dilated screen displacement in pixels, in XY. Python/C pixel decoding.
Motion2dXYZDilated	RGBA	float32	(H, W, 4)
Dilated XY screen displacement in pixels and Z view-depth change; W is 1. Python/C pixel decoding.
MotionWorld	RGBA	float32	(H, W, 4)
World-space displacement in XYZ, in scene units; W is 1. The preview shows XY direction and scales magnitude to the captured 99th percentile. Python/C pixel decoding.
EmissionAndForegroundMask	RGBA	float16	(H, W, 4)
RGB surface emission and W foreground mask. Python/C pixel decoding.
Roughness	RGBA	uint8	(H, W, 4)
Material roughness in the blue channel, stored as a UNORM byte. The preview displays blue as grayscale. Python/C pixel decoding.
DisocclusionMask	Scalar	float16	(H, W, 1)
Disocclusion mask in [0, 1]. The preview displays it directly as grayscale. Python/C pixel decoding.
ViewspacePosition	RGBA	float32	(H, W, 4)
View-space position in XYZ, in scene units. The preview rescales each coordinate. Python/C pixel decoding.
XRDepth	Scalar	float32	(H, W, 1)
Normalized reversed clip depth for XR composition. Python/C pixel decoding.
SelectionState	Four coverage samples	uint16	(H, W, 4)
Selected-surface classification at output resolution. Each channel stores group in bits 0–7 (zero means background) and visibility in bit 8 (1 visible, 0 hidden); bits 9–15 are zero. Channels are adaptive coverage samples, not RGBA colors. Only the nearest selected surface is represented; objects sharing a group form one mask. Requesting this output replaces the built-in selection overlay for this RenderProduct. XR foveation unwarp is unsupported: the output is cleared and a warning is logged. Disable unwarp to read selection in warped coordinates. See the selection SPG example for custom outlines, fills, and hidden-surface cross-hatching.
NormalSD	XYZA	float32	(H, W, 4)
World-space surface normals. Python/C decoding.
DepthSD	Z	float32	(H, W, 1)
Deprecated and still delivered; use DistanceToImagePlaneSD or DistanceToCameraSD instead. For a perspective camera, unitless raster depth mapped from 1 (near clip plane) to 0 (far clip plane); an orthographic camera reverses the mapping, and a fisheye or other lens projection stores the distance from the camera. A pixel whose ray hit nothing holds the far-clip value, which is 0 for a perspective camera. Python/C decoding.
DistanceToCameraSD	Z	float32	(H, W, 1)
Euclidean distance from the camera origin to each surface point, in meters whatever the stage’s metersPerUnit. A pixel whose ray hit nothing holds infinity. Python/C decoding.
DistanceToImagePlaneSD	Z	float32	(H, W, 1)
Perpendicular distance from the image plane to each surface point, in meters whatever the stage’s metersPerUnit. A pixel whose ray hit nothing holds infinity. Python/C decoding.
DiffuseAlbedoSD	RGBA	uint8	(H, W, 4)
Diffuse surface albedo (base color without lighting). Python/C decoding.
Camera3dPositionSD	XYZA	float32	(H, W, 4)
Camera-space 3D position of each visible surface point, in scene units, with W = 1. A pixel whose ray hit nothing has W = 0 and holds a point far along the ray, so test W rather than the position magnitude to find background; a min/max-stretched preview of the raw values comes out flat. Python/C decoding.
TargetMotionSD	XYZW	float32	(H, W, 4)
Backward screen-space displacement in pixels: add XY to the current pixel position to find its previous-frame position. Z is the previous-frame occlusion flag; W indicates a geometric hit within the occlusion-check range. Invalid correspondences are (-1, -1, -1, -1). Python/C decoding.
SemanticSegmentation	ID	uint32	(H, W, 1)
Per-pixel semantic ID. Decode SemanticIdMap to map each integer ID to its semantic label string. Class index 1 marks unlabeled geometry. Class indices occupy the first word of the four-word class ID in SemanticIdMap, so the matching class tuple is (1, 0, 0, 0). Python/C decoding.
SemanticIdMap	IdentifierMap	Named tensors	id (N, 4), label_offsets (N+1,), label_data (B,)
Semantic IDs and UTF-8 labels as uint32 IDs, uint64 offsets and uint8 label bytes. Row order, sparse empty rows and all four ID words are preserved. Python/C decoding.
SemanticInstanceSegmentation	ID[4]	uint32	(H, W, 4)
Stable ID of the semantic prim associated with each pixel. Maintain a prim-path lookup from StableIdMapDeltas. Background is all zeros; unlabeled hits are (0, 0, 0, 1). Python/C decoding.
StableIdSegmentation	ID[4]	uint32	(H, W, 4)
Stable geometry-instance ID per pixel. The four words form one identifier, including its subset component. Background is all zeros. Python/C decoding.
NonStableInstanceSegmentation	ID	uint32	(H, W, 1)
Runtime semantic-instance index plus 2 per pixel. Zero is background; 1 is an unlabeled hit. These indices can change as scene contents change. Python/C decoding.
BoundingBox2DTightSD	BoundingBox2D	uint32, int32	Named tensors
Per-instance visible-pixel bounds: instance_id, render_tile_id and extent (N,2,2). Reject inverted bounds. Python/C decoding.
BoundingBox2DLooseSD	BoundingBox2D	uint32, int32	Named tensors
Per-instance bounds including occluded geometry: instance_id, render_tile_id and extent (N,2,2). Python/C decoding.
BoundingBox3DSD	BoundingBox3D	uint32, float32	Named tensors
Per-instance local extent (N,2,3), object_to_world (N,4,4), instance_id and render_tile_id. Python/C decoding.
SemanticBoundingBox2DTight	SemanticBoundingBox2D	uint32, int32	Named tensors
Visible-pixel bounds grouped by semantic instance: id, semantic_id and extent (N,2,2). Consume only valid_count[0] rows. Python/C decoding.
SemanticBoundingBox2DLoose	SemanticBoundingBox2D	uint32, int32	Named tensors
Bounds including occluded geometry grouped by semantic instance: id, semantic_id and extent (N,2,2). Consume only valid_count[0] rows. Python/C decoding.
SemanticBoundingBox3D	SemanticBoundingBox3D	uint32, float32	Named tensors
Local extent (N,2,3), object_to_world (N,4,4), id and semantic_id. Consume only valid_count[0] rows. Python/C decoding.
OcclusionSD	Occlusion	uint32, float32	Named tensors
instance_id, render_tile_id, and occlusion_ratio have shape (N,). Ratios range from 0 to 1; -1 means no sampled hits. Python/C decoding.
InstanceIdTokenMapSD	InstanceIdTokenMap	uint32, uint64	Named tensors + CPU params
instance_id, prim_token, prototype_index, and instance_index have shape (N,). CPU scalar params provide min_instance_id, max_instance_id, and render_tile_count. Tensor rows contain no header. Python/C decoding.
InstanceMap	InstanceMap	uint32	semantic_instance_index (N,)
Runtime-instance-to-semantic-instance lookup. 0xFFFFFFFF means no semantic parent. Indices address semantic-instance records, not SemanticIdMap label IDs. Python/C decoding.
StableIdMap	IdentifierMap	Named tensors	id (N, 4), label_offsets (N+1,), label_data (B,)
Stable IDs and USD prim paths as uint32 IDs, uint64 offsets and uint8 UTF-8 bytes. Repeated IDs retain their separate rows. Python/C decoding.
StableIdMapDeltas	IdentifierMap	Named tensors	id (N, 4), label_offsets (N+1,), label_data (B,)
New stable-ID/prim-path entries since the last generated delta, using the named identifier-map tensors. The first frame includes scene-creation entries; an empty delta has zero rows and label_offsets = [0]. Python/C decoding.
SemanticIdMapDeltas	IdentifierMap	Named tensors	id (N, 4), label_offsets (N+1,), label_data (B,)
New semantic-ID/label entries since the last generated delta, using the same named tensors as SemanticIdMap. An empty delta has zero rows and label_offsets = [0]. Python/C decoding.
CameraParams	CameraParams	Named CPU params	No tensors
View and projection matrices, scene-unit scale, resolution, clipping range, camera-model name and lens fields as named CPU parameters, for every requested mapping device. Python/C decoding.
StableIdSemanticIdMap	StableIdSemanticIdMap	uint32	Named tensors
stable_id and semantic_id both have shape (N,4). Preserve all four words and repeated rows when joining semantic instances to their labels. Python/C decoding.

The LPE row uses expression source names with sourceType = "lpe". Other rows use the default raw source type.

Minimal

sourceName	Format	Type	Shape	Example
LdrColor	RGBA	uint8	(H, W, 4)
Simple-shaded color in sRGB space. Standard final image output suitable for display or saving as PNG/JPEG.
NormalSD	XYZA	float32	(H, W, 4)
World-space surface normals. Python/C decoding.
DepthSD	Z	float32	(H, W, 1)
Deprecated and still delivered; use DistanceToImagePlaneSD or DistanceToCameraSD instead. For a perspective camera, unitless raster depth mapped from 1 (near clip plane) to 0 (far clip plane); an orthographic camera reverses the mapping, and a fisheye or other lens projection stores the distance from the camera. A pixel whose ray hit nothing holds the far-clip value, which is 0 for a perspective camera. Python/C decoding.
Depth / XRDepth	Z	float32	(H, W, 1)
Unitless reversed clip depth, from 1 (near clip plane) to 0 (far clip plane). Both names expose the same depth values as DepthSD in Minimal mode. Python/C decoding.
DistanceToCameraSD	Z	float32	(H, W, 1)
Euclidean distance from the camera origin to each surface point, in meters whatever the stage’s metersPerUnit. A pixel whose ray hit nothing holds infinity. Python/C decoding.
DistanceToImagePlaneSD	Z	float32	(H, W, 1)
Perpendicular distance from the image plane to each surface point, in meters whatever the stage’s metersPerUnit. A pixel whose ray hit nothing holds infinity. Python/C decoding.
DiffuseAlbedoSD	RGBA	uint8	(H, W, 4)
Diffuse surface albedo (base color without lighting). Python/C decoding.
Camera3dPositionSD	XYZA	float32	(H, W, 4)
Camera-space 3D position of each visible surface point, in scene units, with W = 1. A pixel whose ray hit nothing has W = 0 and holds a point far along the ray, so test W rather than the position magnitude to find background; a min/max-stretched preview of the raw values comes out flat. Python/C decoding.
TargetMotionSD	XYZW	float32	(H, W, 4)
Backward screen-space displacement in pixels: add XY to the current pixel position to find its previous-frame position. Z is the previous-frame occlusion flag; W indicates a geometric hit within the occlusion-check range. Invalid correspondences are (-1, -1, -1, -1). Python/C decoding.
SemanticSegmentation	ID	uint32	(H, W, 1)
Per-pixel semantic ID. Decode SemanticIdMap to map each integer ID to its semantic label string. Class index 1 marks unlabeled geometry. Class indices occupy the first word of the four-word class ID in SemanticIdMap, so the matching class tuple is (1, 0, 0, 0). Python/C decoding.
SemanticIdMap	IdentifierMap	Named tensors	id (N, 4), label_offsets (N+1,), label_data (B,)
Semantic IDs and UTF-8 labels as uint32 IDs, uint64 offsets and uint8 label bytes. Row order, sparse empty rows and all four ID words are preserved. Python/C decoding.
SemanticInstanceSegmentation	ID[4]	uint32	(H, W, 4)
Stable ID of the semantic prim associated with each pixel. Maintain a prim-path lookup from StableIdMapDeltas. Background is all zeros; unlabeled hits are (0, 0, 0, 1). Python/C decoding.
StableIdSegmentation	ID[4]	uint32	(H, W, 4)
Stable geometry-instance ID per pixel. The four words form one identifier, including its subset component. Background is all zeros. Python/C decoding.
NonStableInstanceSegmentation	ID	uint32	(H, W, 1)
Runtime semantic-instance index plus 2 per pixel. Zero is background; 1 is an unlabeled hit. These indices can change as scene contents change. Python/C decoding.
BoundingBox2DTightSD	BoundingBox2D	uint32, int32	Named tensors
Per-instance visible-pixel bounds: instance_id, render_tile_id and extent (N,2,2). Reject inverted bounds. Python/C decoding.
BoundingBox2DLooseSD	BoundingBox2D	uint32, int32	Named tensors
Per-instance bounds including occluded geometry: instance_id, render_tile_id and extent (N,2,2). Python/C decoding.
BoundingBox3DSD	BoundingBox3D	uint32, float32	Named tensors
Per-instance local extent (N,2,3), object_to_world (N,4,4), instance_id and render_tile_id. Python/C decoding.
SemanticBoundingBox2DTight	SemanticBoundingBox2D	uint32, int32	Named tensors
Visible-pixel bounds grouped by semantic instance: id, semantic_id and extent (N,2,2). Consume only valid_count[0] rows. Python/C decoding.
SemanticBoundingBox2DLoose	SemanticBoundingBox2D	uint32, int32	Named tensors
Bounds including occluded geometry grouped by semantic instance: id, semantic_id and extent (N,2,2). Consume only valid_count[0] rows. Python/C decoding.
SemanticBoundingBox3D	SemanticBoundingBox3D	uint32, float32	Named tensors
Local extent (N,2,3), object_to_world (N,4,4), id and semantic_id. Consume only valid_count[0] rows. Python/C decoding.
OcclusionSD	Occlusion	uint32, float32	Named tensors
instance_id, render_tile_id, and occlusion_ratio have shape (N,). Ratios range from 0 to 1; -1 means no sampled hits. Python/C decoding.
InstanceIdTokenMapSD	InstanceIdTokenMap	uint32, uint64	Named tensors + CPU params
instance_id, prim_token, prototype_index, and instance_index have shape (N,). CPU scalar params provide min_instance_id, max_instance_id, and render_tile_count. Tensor rows contain no header. Python/C decoding.
InstanceMap	InstanceMap	uint32	semantic_instance_index (N,)
Runtime-instance-to-semantic-instance lookup. 0xFFFFFFFF means no semantic parent. Indices address semantic-instance records, not SemanticIdMap label IDs. Python/C decoding.
StableIdMap	IdentifierMap	Named tensors	id (N, 4), label_offsets (N+1,), label_data (B,)
Stable IDs and USD prim paths as uint32 IDs, uint64 offsets and uint8 UTF-8 bytes. Repeated IDs retain their separate rows. Python/C decoding.
StableIdMapDeltas	IdentifierMap	Named tensors	id (N, 4), label_offsets (N+1,), label_data (B,)
New stable-ID/prim-path entries since the last generated delta, using the named identifier-map tensors. The first frame includes scene-creation entries; an empty delta has zero rows and label_offsets = [0]. Python/C decoding.
SemanticIdMapDeltas	IdentifierMap	Named tensors	id (N, 4), label_offsets (N+1,), label_data (B,)
New semantic-ID/label entries since the last generated delta, using the same named tensors as SemanticIdMap. An empty delta has zero rows and label_offsets = [0]. Python/C decoding.
CameraParams	CameraParams	Named CPU params	No tensors
View and projection matrices, scene-unit scale, resolution, clipping range, camera-model name and lens fields as named CPU parameters, for every requested mapping device. Python/C decoding.
StableIdSemanticIdMap	StableIdSemanticIdMap	uint32	Named tensors
stable_id and semantic_id both have shape (N,4). Preserve all four words and repeated rows when joining semantic instances to their labels. Python/C decoding.

See Decoding camera outputs for tested Python and C API decoding examples.

Debug Views

Debug views replace the RenderProduct’s beauty output, the shaded color image that normally lands in LdrColor and HdrColor, with diagnostic geometry data or heat maps. Most of them visualize the G-buffer: the per-pixel geometry attributes such as depth, normals, texture coordinates and identifiers that the renderer resolves before shading. Both terms are in the Glossary. Debug views need Real-Time Path Tracing or Path Tracing; Minimal does not serve them. Select a view by authoring its token on the RenderProduct:

string omni:rtx:debug:view:target = "normal"

Token	Description	Example
depth	G-buffer depth multiplied by 1000 and displayed as grayscale. omni:rtx:debug:view:scaling rescales the view; the example image uses 0.005, which fits a scene authored in meters.
worldPosition	World-space position wrapped into a repeating diagnostic color.
barycentrics	Triangle barycentric coordinates.
texcoord0	Fractional values from the first texture-coordinate set.
texcoord1	Fractional values from the second texture-coordinate set.
triangleNormal	Geometric triangle normal encoded into the 0–1 color range.
normal	Interpolated shading normal encoded into the 0–1 color range.
tangentu	First shading tangent encoded into the 0–1 color range.
tangentv	Second shading tangent encoded into the 0–1 color range.
instanceId	Per-pixel instance identifier color.
stableId	Per-pixel stable identifier color. Stable identifiers exist only while a RenderProduct in the same update requests a stable-ID output (StableIdMap, StableIdMapDeltas, StableIdSegmentation, StableIdSemanticIdMap or SemanticInstanceSegmentation); the example image was captured with StableIdSegmentation requested alongside.
nonVisualMaterialId	Per-pixel non-visual material identifier color.
targetMotion	Per-pixel world-space motion, shown with the robots moving.
targetMotionArrows	World-space motion visualized as arrows, shown with the robots moving.
wire	Triangle edges derived from barycentric coordinates.
anyHitCountHeatMap	Color-mapped count of material any-hit calls associated with each pixel, including rejected candidates. Only supported in Real-Time Path Tracing.

Debug views do not create dedicated RenderVars. They replace the beauty destination:

Requested color RenderVars	Debug-view destination
LdrColor and HdrColor	HdrColor. LdrColor remains the regular beauty image.
HdrColor only	HdrColor.
LdrColor only	LdrColor.

Debug views arrive already converted to display colors, so convert a view read from HdrColor to 8-bit directly; applying HDR tonemapping would change the diagnostic colors. RenderVars such as NormalSD and DepthSD remain independent sensor outputs and are not changed by omni:rtx:debug:view:target.

See Python: Debug Views for a runnable example that renders and displays all of these views.

Related

• Decoding camera outputs

• Output Configuration

• Sensor Outputs