Experimental DLSS Neural Rendering:
DLSS5NeuralRendering, displayed as “Experimental DLSS Neural Rendering,” applies an experimental neural-rendering pass to an image or frame sequence through an external DLSS runtime bridge. It provides controls for balancing the rendering effect, style influence, and preservation of local image structure, allowing you to explore different treatments without rebuilding your workflow. For sequences, it can also pass paired depth and motion guidance to the runtime, supplying information about scene distance and how objects move between frames. The node supports selective application of the finished effect and can request automatic masking from the runtime. This is a neural-rendering stage rather than a standalone super-resolution or frame-generation stage; the class does not expose an upscaling factor or frame-interpolation controls. Its results and availability depend on the configured external runtime, and the first use may download a required model. Treat it as an experimental rendering tool rather than a guarantee of a particular visual improvement or a specific NVIDIA-supported integration.
Experimental DLSS Neural Rendering Input Parameters:
image
Required input of type IMAGE. Supplies the image or batch of frames to process; for sequence work, provide frames in their intended temporal order so the runtime and any accompanying motion guidance receive a meaningful sequence. The node copies the image data to CPU memory in 32-bit floating-point format and passes it to the external rendering process using temporary files. Consequently, larger images and longer sequences increase data-transfer, memory, and temporary-storage requirements. The original image data is also used for blending when an effect mask is supplied. No default image, numerical range, or explicit resolution limit is declared by this interface.
style
Required selection with exactly three options: ``0
- neutral
,1, and2. Chooses the style identifier sent to the neural-rendering runtime. Option0 - neutralis documented as the neutral baseline, making it a useful starting point for evaluating the rendering effect before trying other styles. The source does not describe the visual appearance of styles1and2, so you should compare them on representative images rather than assume they correspond to particular artistic looks. No explicit default is declared;0 - neutral`` is the first listed option.
style_strength
Required FLOAT with a minimum of 0.0, maximum of 1.0, default of 1.0, and adjustment step of 0.01. Controls the amount of the selected style requested from the runtime. Lower values reduce the requested style influence, while higher values request more of it. This is separate from overall rendering intensity, so it lets you adjust the selected treatment without using the same control for the entire neural-rendering effect. The value is passed directly to the runtime; the node does not define a mathematical blending rule for this setting or guarantee that 0.0 bypasses processing.
intensity
Required FLOAT with a minimum of 0.0, maximum of 1.0, default of 1.0, and adjustment step of 0.01. Sets the overall neural-rendering effect intensity requested from the runtime. Use this as the main control when you want the complete treatment to be subtler or stronger, rather than only changing the contribution of a selected style. The node forwards this value to the external process without implementing its own intensity blend, so the precise visual response depends on the runtime. In particular, setting it to 0.0 does not skip runtime execution in this class.
local_structure
Required FLOAT with a minimum of 0.0, maximum of 2.0, default of 1.0, and adjustment step of 0.01. Specifies the local structure preservation strength sent to the runtime. In practical terms, this setting concerns how strongly the rendering process is asked to retain nearby shapes and image details while applying its treatment. Higher values request stronger preservation, while lower values relax that request. The exact behavior is runtime-defined, so compare important features such as outlines, facial details, and textured areas when adjusting it rather than assuming that higher values always produce a better result.
skin_structure
Required FLOAT with a minimum of -1.0, maximum of 2.0, default of -1.0, and adjustment step of 0.01. Controls the skin-structure setting passed to the runtime. The special value -1.0 delegates the choice to the runtime, while non-negative values explicitly override it. This makes -1.0 the appropriate baseline when you do not want to impose your own skin-structure preference. For portrait work, compare explicit values within 0.0–2.0 against that baseline to assess their effect on skin detail and facial appearance. Although the interface permits fractional negative values between -1.0 and 0.0, their meaning is not documented; use -1.0 or a non-negative value for documented behavior.
auto_mask
Required BOOLEAN with a default of True; available values are True and False. Requests automatic masking from the external runtime. The source does not describe which areas this automatic masking identifies or how it is calculated, so it should not be assumed to mean a specific face, skin, or subject-selection method. This request is independent of the optional effect mask: automatic masking influences runtime processing, whereas an effect mask is applied afterward to blend the rendered result with the original image. Disable it when you want to compare the runtime’s behavior without automatic masking, but evaluate the actual result because its detailed operation is runtime-dependent.
depth_inverted
Required BOOLEAN with a default of False; available values are True and False. Tells the runtime whether the supplied depth guide uses an inverted near/far convention. Depth guides describe relative scene distance, but different depth-producing tools can represent nearby and distant surfaces in opposite ways. Enable this only when the supplied guide’s near/far ordering needs inversion for the runtime. The node passes the flag to the runtime rather than modifying the depth image itself. Without a depth guide, there is no supplied depth information for this setting to reinterpret.
effect_mask
Optional input of type MASK, with no default mask. Limits where the final neural-rendered image replaces the original image. A mask value of 0 preserves the original, a value of 1 uses the rendered result, and intermediate values blend the two; values are clamped to the 0–1 range before blending. The node resizes the mask to the rendered image dimensions using bilinear interpolation, which can produce soft transitions along boundaries. A single mask is repeated across all output frames when processing a multi-frame result; for changing regions across a sequence, supply a compatible per-frame mask batch. This mask is applied after the external rendering process, so it does not restrict which pixels the runtime processes or necessarily reduce rendering cost.
depth
Optional input of type IMAGE, with no default guide. Provides depth information describing scene distance for temporal guidance. You must connect it together with motion_vectors or leave both disconnected; supplying depth alone causes an explicit validation error. When both guides are provided, the node passes them through a guide-saving helper using the source image’s height and width as the target guide size, then supplies the saved guides to the external process. Use depth that corresponds to the same images and frame order as the main sequence. The complete accepted depth layout and normalization rules are not shown in this class, so prefer a compatible depth-guide node from the same extension rather than assuming any grayscale image will be interpreted correctly.
motion_vectors
Optional input of type IMAGE, with no default guide. Supplies a current-to-previous motion guide, describing how positions in the current frame relate to the preceding frame. Together with depth, this provides the runtime with temporal guidance for frame sequences. It must be connected together with depth; supplying motion vectors alone raises a validation error. Direction matters: a previous-to-current flow field is not the documented guide direction. Use motion guidance generated for the same frame sequence and ordering as the image input, preferably through compatible guide nodes from this extension. This class does not show the full channel encoding or motion-value units, so an ordinary flow visualization should not automatically be treated as a valid motion-vector input.
native_mode
Optional COMBO[STRING] input that selects the native execution mode used by the external-process helper. The function defaults to the first entry in NATIVE_MODES, and the selected mode is also passed to the helper that appends native execution information to the report. The supplied source references NATIVE_MODE_INPUT but does not include its definition, so the exact option labels and default label cannot be established from this excerpt. Use the options shown in your installed node and retain the default unless your runtime setup documentation directs you to choose another mode; specific performance or compatibility differences cannot be inferred from this class alone.
Experimental DLSS Neural Rendering Output Parameters:
image
Output of type IMAGE containing the neural-rendered image or frame batch loaded from the external process. If an effect mask was supplied, this is the final blend of the original image and the runtime’s rendered result, not the unmasked rendering. The node moves the output tensor back to the same device as the original image input before returning it, allowing you to connect it to subsequent image-processing or saving nodes. This class does not implement an explicit resizing or frame-generation stage; the returned image dimensions and frame count are recorded in the accompanying report.
runtime_report
Output of type STRING containing a readable processing summary followed by native execution details supplied by a helper. The summary starts with DLSS-NR 310.8 Feature 18 and includes the returned width and height, frame count, a temporal label, and whether paired guides were supplied. The temporal label is yes when the result contains more than one frame and still-baseline otherwise; this is calculated from frame count and is not proof that temporal processing succeeded or that flicker was eliminated. The guide label is depth+motion when both guides were connected and none otherwise. Use this report to confirm what the node returned and what guidance was supplied, and retain it when investigating runtime problems.
Experimental DLSS Neural Rendering Usage Tips:
- Start with
styleset to0 - neutral,local_structureat 1.0, andskin_structureat -1.0. Keep the other controls at their defaults for your first comparison, then change one setting at a time. - For a subtler treatment, lower
intensityin small increments. If the overall rendering is useful but the selected style is too pronounced, adjuststyle_strengthseparately. - For portraits, compare the default
skin_structurevalue of -1.0 with non-negative overrides on the same frame. Avoid undocumented fractional negative values, and inspect facial details rather than judging only the overall image. - For guided sequence processing, connect both
depthandmotion_vectorsand make sure both guides correspond to the same frame order asimage. The motion guide must use the documented current-to-previous direction. - Leave
depth_inverteddisabled unless you know your depth guide’s near/far convention requires inversion. Do not use it as a general strength or contrast control. - Use
effect_maskfor selective results, such as preserving a background while applying the rendering to a subject. Soft mask edges provide blended transitions, and a moving subject generally needs a matching sequence of masks rather than one static mask. - Compare
auto_maskenabled and disabled when diagnosing unexpected selective effects. A supplied effect mask still limits the final result independently of this runtime setting. - Test a short, lower-resolution sequence before processing a long batch. This class writes the full input batch to temporary storage and has no exposed chunk-size control, so large jobs can increase memory and disk demands.
- Allow for a possible model download on first use, and verify your bridge setup with the extension’s runtime setup or status tools. Check
runtime_reportafter a successful test before committing to a larger job.
Experimental DLSS Neural Rendering Common Errors and Solutions:
depth and motion_vectors must be connected together
- Explanation: This exact validation error occurs when only one of the two optional guidance inputs is connected. The node accepts paired depth and motion guidance or no guidance.
- Solution: Connect both
depthandmotion_vectors, or disconnect both. For sequences, generate both from the same ordered frames.
DLSS 5 failed:
- Explanation: This exact prefix is raised when the external process returns a nonzero exit code. The node appends the process’s standard output and error text, which contain the more specific cause; the prefix alone does not identify the problem.
- Solution: Read the full message after the prefix. Check the extension’s runtime status, verify the configured runtime components, and try a small image without optional guides to isolate setup problems from guide-related problems.
Runtime path, plugin, or snippet errors — message varies
- Explanation: The node obtains paths for an external Python executable, plugin, and snippet before rendering. Missing files or an incomplete runtime installation can prevent the process from starting or completing; the exact messages are defined outside this class.
- Solution: Run the extension’s runtime setup and status tools, confirm that the configured files exist and are accessible, and follow any specific diagnostic instructions before retrying.
First-use model download failure — message varies
- Explanation: The node description states that first use may download a required model. If that download is blocked or cannot be saved, rendering may fail before a result is produced.
- Solution: Check network access, available disk space, and write permissions for the runtime’s model location. Retry after resolving the reported download issue, using the full runtime message to identify the affected resource.
Runtime timeout — message varies
- Explanation: The external process is called with a configured runtime timeout. A slow initialization, model download, or demanding job may exceed that limit; the exact timeout error is not shown in this class.
- Solution: Retry with a smaller image or shorter sequence and check whether first-use setup is still occurring. If your installation exposes timeout configuration, follow its documentation rather than looking for a timeout parameter on this node.
Temporary-file read or write failure — message varies
- Explanation: Images, optional guides, and the returned result are exchanged through temporary NumPy files. Insufficient disk space, inaccessible temporary directories, or file permissions can interrupt this exchange.
- Solution: Free space in the configured runtime temporary location, verify that ComfyUI and the external process can access it, and retry a smaller job. If a result file is missing, also inspect the external runtime logs because the renderer may have failed to create it.
Out-of-memory error — message varies
- Explanation: The node creates a CPU floating-point copy of the entire input batch, and the external renderer may have additional memory requirements. Large frames, long batches, and guides can increase memory usage.
- Solution: Reduce resolution or batch length, close other memory-intensive applications, and test again. If you divide a sequence manually, keep each segment’s frames and guides aligned and inspect continuity at segment boundaries.
Guide shape or format error — message varies
- Explanation: Incompatible guide dimensions, frame counts, channel layouts, or motion encodings may cause failures in guide preparation or in the external runtime. The exact validation rules and messages are not included in this class.
- Solution: Regenerate guides with compatible nodes from the extension, confirm that they match the image sequence, and verify current-to-previous motion direction. Do not substitute a colorized motion preview for encoded motion data unless it is explicitly supported.
Mask batch or tensor-size mismatch — message varies
- Explanation: A single mask can be expanded across a multi-frame result, but this class does not reconcile arbitrary mismatches between multiple masks and the output frame count. Incompatible dimensions can fail during resizing or final blending.
- Solution: Supply a standard MASK with either one mask for the full batch or one matching mask per output frame. Temporarily disconnect
effect_maskto confirm whether the error occurs during final blending rather than rendering.
Invalid style value — message varies
- Explanation: The node converts the first space-separated part of
styleto an integer. Normal dropdown choices are valid, but a malformed value supplied through a modified workflow or API request can fail this conversion. - Solution: Restore one of the exact supported choices:
0 - neutral,1, or2. Avoid replacing the selection with a descriptive style name.
