TS DLSS Upscaler:
TS DLSS Upscaler uses NVIDIA DLSS 5 Neural Rendering to re-render a picture or a sequence of video frames at the same size or a higher resolution. It is designed to recover detail and adjust the rendered look while giving you control over how strongly the image changes. It supports a built-in v9 engine and an optional, separately installed v21 application. The node runs on Windows with an NVIDIA RTX GPU; it is not a general-purpose upscaler for other platforms.
TS DLSS Upscaler Input Parameters:
images
The picture or batch of frames to process, supplied as an IMAGE tensor shaped [B,H,W,C]. Both width and height must be at least 64 pixels. A batch can represent consecutive video frames when temporal processing is enabled, or unrelated pictures when it is disabled. RGB channels are rendered by the network; if the image has an alpha channel, the node carries that channel to the output size separately.
upscaling_factor
Sets the requested output size. A 1× selection keeps the original dimensions while still re-rendering the image; higher selections request an enlarged result. Output dimensions are capped at 7680×4320. The available options and default come from UPSCALING_LABELS and DEFAULT_FACTOR_LABEL, whose values are not present in the supplied source, so their exact labels cannot be listed here.
dlss_model_preset
A compatibility control for older workflows. Its value is ignored by both engines: changing it does not select a model or switch between v9 and v21. The default is M; the complete option list comes from LEGACY_MODEL_PRESETS, which is not included in the supplied source. Use engine to choose the backend instead.
nr_style
Chooses the neural-rendering look: Default, Natural, or Cinematic. The default is Default. This changes the style requested from the runtime rather than the output dimensions; compare the looks on a representative frame before processing a long batch.
nr_intensity
Controls the overall strength of neural rendering in both engines. The range is 0.0–1.0, the default is 1.0, and the step is 0.01. Lower values reduce changes across the image, including changes to faces and skin, making this a useful first adjustment when the result looks too heavily processed.
local_tone_strength
Controls how much local contrast the network recovers; lower values produce a softer tonal result. The input offers 0.0–2.0 in steps of 0.01, with a default of 1.0. v9 accepts that full range, but v21 accepts only 0.0–1.0 and raises a validation error for higher values.
local_structure_strength
Controls the recovery of local detail and texture; lower values make texture enhancement gentler. The input offers 0.0–2.0 in steps of 0.01, with a default of 1.0. v9 accepts the full range, while v21 requires 0.0–1.0 and raises an error above 1.0.
skin_structure_strength
Adds skin detail when automatic_mask is enabled. The input range is −1.0–2.0 in steps of 0.01, and the default is 0.0. Keeping it at 0 avoids extra skin texture, but does not prevent other processing from changing a face. High values can exaggerate pores; the source specifically warns that 2.0 in v9 can produce honeycomb-like mottling. v21 accepts only 0.0–1.0 and errors outside that range.
automatic_mask
Lets the runtime decide where neural rendering is applied. Options are on or off; the default is on. It is needed for the skin-structure control in both engines and for face_skin_protection in v9. Turn it on when you want those controls to take effect.
temporal
Treats a multi-image batch as consecutive video frames. Options are on or off; the default is on. With v9, frames share temporal history and that history resets at scene cuts; with v21, the application processes the batch as one video using its own temporal policy. Turn this off for unrelated pictures so they are not treated as a continuous sequence. It has no effect on a single image.
source_curve
Describes what the incoming pixel values mean. The default is the SDR label from TRANSFER_LABELS; the exact option labels are not included in the supplied source. The network works with display-referred SDR, so the node converts selected log, HDR, or linear sources before rendering and applies the corresponding inverse curve afterward. Choose the curve that matches your source rather than using it as a creative look control.
dither
A compatibility control with no effect in either engine. Options are on or off; the default is on. v9 uses float32 frame transport, while v21 uses its own PNG16 or video transport, so changing this setting does not alter the rendered result.
download_if_missing
Controls whether a missing v9 runtime can be fetched into models/DLSS. Options are on or off; the default is on. The first download is approximately 112 MB. Turning it off prevents that missing-runtime download, though the source notes that an already installed older v5 runtime from the same project may still be updated. This control never downloads v21: you must install the complete v21 application separately.
nr_passes
Sets how many times the network evaluates the image. The range is 1–4, and the default is 1. Each additional pass costs another full evaluation, so begin with one pass and increase it only when a comparison justifies the extra processing time. Both engines use this setting.
nr_color_strength
Controls how much of the network’s colour reaches the v9 result. The range is 0.0–1.0 in steps of 0.01, and the default is 1.0; lowering it keeps more of the source’s colour. In v21, this setting must remain at 1.0—another value causes an error rather than providing colour preservation.
tone_preservation
In v9, preserves more of the source’s tonality against the network’s tonal changes as you raise it. The range is 0.0–1.0 in steps of 0.01, and the default is 0.0. v21 does not expose this control and requires it to stay at 0.0.
face_skin_protection
In v9, adds restraint to changes on faces and requires automatic_mask to be on. The range is 0.0–1.0 in steps of 0.01, and the default is 0.0. v21 requires it to remain at 0.0; for gentler v21 results, lower nr_intensity and keep skin_structure_strength at 0.0.
grain_preservation
In v9, keeps more of the source’s grain instead of letting the network clean it away, which can help with film scans. The range is 0.0–1.0 in steps of 0.01, and the default is 0.0. v21 requires this control to stay at 0.0.
shimmer_suppression
In v9, controls a GPU temporal stabiliser that steadies detail between frames. The range is 0.0–1.0 in steps of 0.01, and the default is 0.70. It is used only when temporal is on and the batch contains more than one frame; for a still image, the node renders with a value of 0. In v21 this value is ignored because the external application controls temporal stabilisation.
engine
Selects v9 (built-in) or v21 (external CLI); the default is v9 (built-in). v9 uses the built-in integration and its runtime in models/DLSS. To use v21, download the complete official Visual Enhancer v21.0 archive, extract it into runtime/visual-enhancer/v21 inside the node pack with VE_CLI.exe directly in that folder, and then select v21. The node does not download the v21 application.
v21_directory
An optional folder setting used only by v21; the default is an empty string. Leave it blank to use runtime/visual-enhancer/v21 inside the node pack, or specify a relative location inside the pack. Absolute paths and paths that use ../ to leave the pack are rejected. This setting is ignored when engine is v9.
TS DLSS Upscaler Output Parameters:
images
The processed IMAGE batch, in the same frame order as the input and at the selected output dimensions, subject to the 7680×4320 cap. The result retains the input’s channel count after any required grayscale-to-RGB expansion: RGB is rendered by the selected engine, while a source alpha channel is resized and carried through separately rather than processed by the network. Use this output wherever you need the finished still image or the ordered frames for the next video-processing step.
TS DLSS Upscaler Usage Tips:
- Start with one representative image,
nr_passesat 1, and the default v9 engine. Compare a 1× re-render against an enlarged result to separate changes in rendering style from changes in size. - For unrelated pictures in one batch, turn
temporaloff. For consecutive video frames, leave it on; v9 can then use temporal history andshimmer_suppression. - If faces or textures look overprocessed, lower
nr_intensityfirst. Keepskin_structure_strengthat 0 unless you specifically want more skin detail, and reducelocal_structure_strengthfor gentler texture. - When using v21, install the complete application before running the node. Keep
nr_color_strengthat 1 andtone_preservation,face_skin_protection, andgrain_preservationat 0; also keep skin, local-tone, and local-structure strengths within 0–1. - Matchsource_curveto the incoming image. If the source is already display-referred SDR, leave it on the SDR selection; selecting a log, HDR, or linear curve for an SDR image requests an unnecessary conversion. - For grainy source material on v9, try raising
grain_preservation. For video detail that shimmers, testshimmer_suppressionon a short sequence before processing the full batch.
TS DLSS Upscaler Common Errors and Solutions:
DLSS 5 Neural Rendering runs through NVIDIA's signed D3D12 runtime, which exists for Windows only.
- Explanation: The node stops on platforms other than Windows.
- Solution: Run the workflow on Windows with a compatible NVIDIA RTX setup.
numpy is required.
- Explanation: The node could not import NumPy, which it needs to prepare and assemble frames.
- Solution: Install NumPy in the Python environment used by ComfyUI, then restart ComfyUI.
'images' must be a batch shaped [B,H,W,C].
- Explanation: The supplied image is missing or is not in the four-dimensional batch format expected by the node.
- Solution: Connect a ComfyUI IMAGE output that provides a batch, even when it contains only one picture.
DLSS needs at least 64x64 pixels; this batch is …
- Explanation: The input width or height is below the minimum of 64 pixels.
- Solution: Supply a larger source image or resize it so both dimensions are at least 64 pixels before this node.
Unknown engine: …
- Explanation: The engine value does not match either supported selection.
- Solution: Select
v9 (built-in)orv21 (external CLI)from the node’s engine menu.
v21 CLI does not expose … Reset those v9-only controls to their defaults or choose v9.
- Explanation: At least one v9-only colour, tone, face-protection, or grain control was changed while v21 is selected.
- Solution: Set
nr_color_strengthto 1.0 andtone_preservation,face_skin_protection, andgrain_preservationto 0.0, or switch to v9 to use those controls.
v21 requires skin/tone/structure between 0 and 1; use v9 for the wider legacy ranges.
- Explanation: v21 received a value outside 0–1 for
skin_structure_strength,local_tone_strength, orlocal_structure_strength. - Solution: Bring all three values into the 0–1 range, or choose v9 if you need its wider ranges.
Missing v9 runtime or runtime download failure
- Explanation: The built-in engine needs its runtime in
models/DLSS. A first-time run may need to download it, and a disabled download or failed network request can prevent startup; the exact error wording depends on the runtime installer. - Solution: If you permit the download, turn
download_if_missingon and retry with network access. Otherwise, provide the required v9 runtime locally before running the node.
Missing or invalid v21 application directory
- Explanation: v21 is an external application, not part of the v9 runtime download. An incomplete installation, a missing
VE_CLI.exe, or a directory path outside the node pack can prevent it from launching; the exact message comes from the external integration. - Solution: Extract the complete official v21.0 archive into the documented folder with
VE_CLI.exedirectly inside it. Leavev21_directoryblank for that location, or use a valid relative path within the pack.
The run finished but no frame was recorded as evaluated — the result may not be DLSS-enhanced.
- Explanation: This is a v9 warning, not a rendering error: the run completed, but the node could not verify from its recorded evidence that a frame was evaluated.
- Solution: Inspect the output before relying on it as DLSS-enhanced, and check the ComfyUI log for the engine’s status and other warnings.
