Save 4 hours! We auto-setup your workflow! Free!

Drop your workflow.json — we handle every dependency, custom node, and model. Just open the link and run.

Auto-Setup Workflow Json (Free) Now!
ComfyUI > Nodes > ComfyUI-RH-DLSS5 > RH DLSS5 Enhance (NR / Upscale)

ComfyUI Node: RH DLSS5 Enhance (NR / Upscale)

Class Name

RH_DLSS5Enhance

Category
image/upscaling
Author
RH-RunningHub (Account age: 210days)
Extension
ComfyUI-RH-DLSS5
Latest Updated
2026-11-11
Github Stars
0.05K

How to Install ComfyUI-RH-DLSS5

Install this extension via the ComfyUI Manager by searching for ComfyUI-RH-DLSS5
  • 1. Click the Manager button in the main menu
  • 2. Select Custom Nodes Manager button
  • 3. Enter ComfyUI-RH-DLSS5 in the search bar
After installation, click the Restart button to restart ComfyUI. Then, manually refresh your browser to clear the cache and access the updated list of nodes.

Visit ComfyUI Online for ready-to-use ComfyUI environment

  • Free trial available
  • 16GB VRAM to 80GB VRAM GPU machines
  • 400+ preloaded models/nodes
  • Freedom to upload custom models/nodes
  • 200+ ready-to-run workflows
  • 100% private workspace with up to 200GB storage
  • Dedicated Support

Run ComfyUI Online

RH DLSS5 Enhance (NR / Upscale) Description

Enhances images and video with NVIDIA DLSS 5 neural rendering and optional 1.5×–3× upscaling.

RH DLSS5 Enhance (NR / Upscale):

RH_DLSS5Enhance applies NVIDIA DLSS 5 Neural Rendering to enhance existing artwork and video, with optional enlargement through the standard DLSS upscaling stage before neural enhancement. It offers native-resolution refinement as well as supported enlargement factors from 1.5× to 3×, making it useful for refining generated images or preparing larger versions of your work. You can adjust the overall rendering style, local tone, reconstructed structure, skin treatment, and how much of the model’s color and brightness changes are retained. For moving footage, it can use motion estimates and information from preceding frames to support temporal consistency; for unrelated pictures, it can reset that history between frames. Enhancement is model-driven, so changes should be reviewed rather than treated as guaranteed recovery of original detail. Windows processing uses an in-process D3D12 bridge, while Linux processing sends frames to a Windows enhancement host running under Wine. The necessary NVIDIA runtime DLLs must be supplied separately; they are not redistributed with the extension.

RH DLSS5 Enhance (NR / Upscale) Input Parameters:

image

An optional IMAGE connection containing one image or a batch of images. Each batch entry is treated as a frame, in its existing order, so a batch can represent either independent artwork or a coherent sequence. Choose the appropriate batch_mode to determine whether processing should retain information between those entries. There is no numeric range or explicit default image. Although this connection is optional, you must connect at least one of image or video; connecting both causes both sources to be processed.

video

An optional VIDEO connection, such as one supplied by LoadVideo. Frames are processed in temporal order, making this suitable for enhancing or enlarging a complete clip. For file-backed video, the node attempts chunked decoding rather than loading the entire clip into memory; if that reader cannot be created, it falls back to full decoding, which can require substantially more memory. There is no numeric range or explicit default video. You must connect this or image, and both can be connected simultaneously.

upscaling_mode

Selects whether to enhance at the original dimensions or enlarge before neural enhancement. The supplied context identifies fixed factors of 1×, 1.5×, 1.724×, 2×, and 3×, corresponding respectively to native/DLAA, Quality, Balanced, Performance, and Ultra Performance modes, as well as automatic 1K, 2K, 4K, and 8K modes. The exact dropdown labels for the fixed factors are not shown because the source references an external _UPSCALE_MODES list; the declared default is that list’s first entry, whose literal value is not included. Automatic modes target short-edge lengths of 1080, 1440, 2160, and 4320 pixels and select the smallest supported factor that reaches the target, preserving the source aspect ratio rather than forcing a fixed canvas. Because factors are discrete, not every source will land exactly on the target. Sources already at or above the target remain at 1× rather than being reduced. Unreachable targets or results beyond the 7680-pixel long-edge and 4320-pixel short-edge envelope produce an error. The bypass style overrides this setting and disables enlargement too.

style

Options: default, natural, cinematic, and off (bypass NR); default: default. This selects the neural rendering treatment: natural stays closer to the source, while cinematic pushes contrast. The source does not further define the visual character of default. Selecting off (bypass NR) skips all enhancement and upscaling and passes frames through unchanged, which is useful for comparison or temporarily disabling processing. A file-backed video is returned directly in this mode, without decoding and rebuilding its frames.

preset

An integer internal render-preset hint passed to the runtime as DLSSNR.Hint.Render.Preset. Minimum: 0; maximum: 9; default: 0; step: 1. These numbers are not documented here as artistic presets or as a quality ladder, so a higher number should not be assumed to improve results. Keep this at 0 unless the runtime you are using explicitly documents another value.

intensity

Controls the strength of the neural enhancement pass. Minimum: 0.0; maximum: 2.0; default: 1.0; step: 0.05. Values below 1.0 blend the treatment toward the source, which can help when the result changes your artwork more than you want. Values above 1.0 usually have no additional effect according to the source tooltip, so increasing this control is not a reliable way to obtain stronger reconstruction. This is distinct from detail, which controls the final composite strength.

tone

Controls local tone-mapping strength through DLSSNR.LocalToneStrength. Minimum: 0.0; maximum: 2.0; default: 1.0; step: 0.05. Local tone mapping concerns brightness relationships within different areas of the picture, rather than simply applying a single brightness change to the whole frame. Use it to adjust the model’s local tonal treatment, and compare shadow and highlight areas carefully. The source does not guarantee a specific visual response at each value.

structure

Controls local detail and structure reconstruction through DLSSNR.LocalStructureStrength. Minimum: 0.0; maximum: 2.0; default: 1.5; step: 0.05. This adjusts the neural model’s treatment of fine structures and local detail, rather than changing output dimensions. Lower it when reconstructed textures or edges look too assertive, and test higher values when you want stronger structural treatment. Because this is learned reconstruction, stronger settings should not be interpreted as more accurate recovery of the original scene.

skin

Controls skin-specific structure strength through DLSSNR.SkinStructureStrength, but only when auto_mask is on. Minimum: -1.0; maximum: 2.0; default: -1.0; step: 0.05. Negative values leave the parameter at the model’s default rather than requesting negative enhancement; nonnegative values provide an explicit override. This lets you adjust skin reconstruction separately from the broader structure treatment when the model identifies skin regions. If auto_mask is off, changing this setting has no effect.

auto_mask

Options: on and off; default: on. Enables the model’s automatic detection of skin regions for skin-specific reconstruction. This also determines whether the skin setting is active. Keep it enabled when you want to tune skin treatment independently, or disable it when you do not want that skin-specific control applied. It is an internal model option, not a connection for supplying or exporting a mask.

motion

Options: auto, nvof, optical_flow, and none; default: auto. Selects how motion between frames is estimated for temporal accumulation, which allows processing to use information across a sequence. auto prefers NVIDIA hardware optical flow when available, then OpenCV DIS optical flow, and finally zero motion if neither is available. nvof requests NVIDIA hardware optical flow but falls back to DIS with a note when necessary. optical_flow explicitly selects OpenCV DIS, while none supplies zero motion. Zero motion does not itself mean that temporal history is reset; use batch_mode to control history handling.

scene_change_threshold

Sets the mean luminance-change threshold used to detect scene cuts and reset temporal history in the optical-flow path. Minimum: 0.01; maximum: 1.0; default: 0.24; step: 0.01. A lower threshold triggers resets for smaller brightness changes, while a higher threshold requires a larger change. Resetting history helps avoid carrying information from one shot into the next, but an overly sensitive setting may also reset during flashes or other genuine brightness changes within a shot. This control is documented as optical-flow-only, not as a universal reset threshold for every processing path.

batch_mode

Options: temporal sequence and still images; default: temporal sequence. temporal sequence preserves history between frames and is appropriate for video or coherent image sequences. still images resets temporal history for each frame, making it the appropriate choice for a batch of unrelated artworks or photographs. Selecting the correct mode prevents independent pictures from being treated as though they were consecutive moments in the same scene.

warmup_frames

An integer warm-up budget reported to the processing worker. Minimum: 0; maximum: 120; default: 0; step: 1. The source explicitly notes that 0 is acceptable because the first frame initializes temporal history through a reset. The supplied code does not establish a guaranteed quality improvement from larger values or precisely describe how every runtime uses this budget. Leave it at 0 unless you have a specific reason to test another value.

keep_audio

Options: on and off; default: on. Controls whether original audio is retained when constructing the VIDEO result. When a source file is available, audio is stream-copied from that file; otherwise, available audio attached to the video components is re-encoded. This does not generate audio for a silent source. There is an important bypass exception: with style set to off (bypass NR), a file-backed VIDEO object is returned unchanged, so this setting does not strip its audio in that path.

backend

Options: auto, linux-wine, and windows-bridge; default: auto. Chooses the processing backend rather than an artistic style or quality level. auto selects the backend appropriate to the operating system. linux-wine streams frames to dlss5nr_host.exe under Wine, while windows-bridge uses an in-process D3D12 bridge on Windows. Leave it on auto for normal use; forcing a backend requires an environment capable of running that backend.

channel_order

Options: auto, RGBA, and BGRA; default: auto. Specifies how to interpret the enhancement result’s color-channel layout. auto checks the first frame for a red/blue swap; explicit RGBA or BGRA overrides that detection. This is a compatibility correction, not a creative color adjustment. If red objects become blue or the overall result has an obvious red/blue reversal, test an explicit layout.

hdr

Options: off and on; default: off. Determines whether the neural model receives linear-light values or the default sRGB-encoded proxy. on uses linear light, matching the reference tool’s --nr-hdr behavior and offering a higher-fidelity tonal path at the cost of a color-space round trip. off retains the default processing path used by existing workflows. This setting describes the model’s processing space; it does not by itself establish HDR video export, HDR metadata, or a particular output bit depth.

global_tone

Controls the runtime’s global tone-strength parameter, DLSSNR.GlobalToneStrength, intended to govern overall brightness changes. Minimum: -1.0; maximum: 2.0; default: -1.0; step: 0.05. The documented sentinel value -1 leaves the model default unchanged; 0 requests that global tone be disabled, and values from 0 to 2 provide a manual override. The provided tooltip explicitly states that this control has no effect on the tested 310.8 runtime, though it may work with other DLL builds. No separate behavior is documented for negative values other than the default sentinel.

detail

Controls how strongly the neural-rendered result is composited over the original frame in linear light. Minimum: 0.0; maximum: 2.0; default: 1.0; step: 0.05. At 0, the composite keeps the original appearance; at 1, it uses the full neural-rendered result; above 1, it amplifies the model’s changes beyond their normal strength, including brightness or denoising changes. Despite its name, this is not solely a fine-detail or sharpness adjustment. Reducing it can make enhancement more restrained, while increasing it can exaggerate unwanted changes. It is not documented as disabling processing or cancelling the selected output size; use the bypass style for that.

color

Controls how much of the model’s color treatment is retained. Minimum: 0.0; maximum: 1.0; default: 1.0; step: 0.05. At 0, the composite preserves the original hue while taking the model’s luminance treatment; at 1, it fully follows the model’s color. Intermediate values provide a compromise. Lower this setting when you like the enhanced brightness or structure but want to protect an established palette, skin hue, or product color from model-induced shifts.

runtime_dir

An optional STRING specifying the folder containing the NVIDIA runtime DLLs. Default: an empty string; no numeric range applies. With no explicit path, the node checks DLSS5_RUNTIME_DIR and then the extension’s runtime/ folder. The tooltip identifies nvngx_dlssnr.dll as the neural-rendering runtime and nvngx_dlss.dll as additionally necessary for enlargement beyond 1×; shared helpers also indicate GPU-specific neural-runtime selection can occur. These DLLs must be supplied separately. Enter the containing folder, not a DLL filename, and use compatible runtime files rather than assuming any similarly named build will work.

wine_prefix

An optional STRING specifying the Linux worker’s Wine prefix, which is the folder containing its configured Windows-compatible environment. Default: an empty string; no numeric range applies. An empty value uses DLSS5_WINEPREFIX first and then ~/.wine. The prefix is validated at runtime and needs DXVK and DXVK-NVAPI installed. This setting is Linux-only and is ignored on Windows, where processing runs natively.

RH DLSS5 Enhance (NR / Upscale) Output Parameters:

images

An IMAGE result containing the processed image batch when image is connected. Batch entries correspond to the supplied images or frames, with dimensions determined by the selected scaling mode. Connect this to a standard SaveImage node to save the result. In bypass mode, the original image tensor is returned unchanged. The shown code initializes this result to no image when no image source is supplied, so do not assume that connecting only a VIDEO will also provide a fully decoded IMAGE batch.

video

A VIDEO result suitable for a standard SaveVideo node. This is the video-oriented result for supplied video, with audio retention governed by keep_audio when the video is rebuilt. The shown bypass implementation also attempts to construct a 24 fps video from an image batch; that should not be taken as a guarantee of identical image-to-video behavior in every processing path because the remaining video-processing source is truncated. A file-backed video bypass returns the original VIDEO object unchanged. If both sources are connected, there is still only one VIDEO socket, not separate video results for each source; the shown bypass path gives the supplied video precedence over a video constructed from images.

status

A STRING containing a human-readable processing summary. For the image-processing path shown, it reports the frame count, original and resulting dimensions, selected backend, channel order, and motion engine, with additional notes for HDR, an explicitly selected nonnegative global-tone value, or automatic scaling when applicable. Bypass messages state that neural rendering was skipped and frames were passed through unprocessed. Use this text to confirm how the node actually ran and to help diagnose unexpected results; it is a summary, not a numerical quality assessment or a substitute for reviewing the artwork.

RH DLSS5 Enhance (NR / Upscale) Usage Tips:

  • For a batch of unrelated artworks, select batch_mode = still images so that every picture starts with fresh temporal history. For video or coherent animation frames, use temporal sequence and start with motion = auto.
  • For a restrained treatment, try style = natural, leave intensity at 1.0 initially, and reduce detail if the overall result is too strong. Lower color if you want to retain more of the original palette.
  • For portraits, keep auto_mask = on and begin with skin = -1 to use the model default. If skin reconstruction needs adjustment, compare explicit skin settings while keeping other controls fixed.
  • Start at native resolution when evaluating tone, structure, or color. Once the treatment looks right, enable enlargement so that changes in pixel dimensions do not complicate your comparison.
  • Choose automatic resolution modes by the source’s short edge, not just familiar video labels. A portrait or square image will not have the same dimensions as a 16:9 image using the same bucket, and fixed scale factors may exceed the nominal target.
  • If an automatic target cannot be reached within 3× enlargement or the output envelope, choose a smaller bucket or a valid fixed factor rather than repeatedly increasing enhancement strength.
  • Use style = off (bypass NR) for an unprocessed comparison. Remember that this disables upscaling as well; detail = 0 is not the same execution shortcut.
  • Keep preset = 0, warmup_frames = 0, backend = auto, and channel_order = auto as your starting configuration. Change compatibility controls only when testing a specific issue.
  • Test hdr = on on a short clip or representative image before committing to a long render. Compare tonal transitions, and do not assume this option alone produces an HDR-encoded deliverable.
  • Prefer file-backed VIDEO connections for long clips so the node can attempt chunked decoding. If the console reports full-decoding fallback, test a shorter clip first and monitor memory use.
  • When checking motion problems, use a short section containing both movement and a cut. Compare auto and optical_flow; adjust scene_change_threshold only in the applicable optical-flow path, lowering it if cuts are not triggering sufficient resets.
  • Save image results through SaveImage and video results through SaveVideo. Review status alongside the saved result to confirm scaling, backend selection, and bypass behavior.

RH DLSS5 Enhance (NR / Upscale) Common Errors and Solutions:

The enhance node needs at least one input: connect image or video.

  • Explanation: This is an exact error from the provided source. All connections are declared optional in the interface, but execution still requires at least one media source.
  • Solution: Connect an IMAGE source, a VIDEO source, or both, then run the workflow again.

Automatic scaling target cannot be reached

  • Explanation: This is a descriptive error category; the exact message is not supplied. Automatic modes fail when the requested short-edge target cannot be reached using the supported factors up to 3×, or when reaching it would exceed the output envelope.
  • Solution: Select a smaller automatic bucket or a supported fixed factor. Check the source dimensions and aspect ratio before requesting another target.

Requested dimensions exceed the supported output envelope

  • Explanation: This is a descriptive validation issue, not a quoted error. The supplied scaling guidance limits output to a long edge of 7680 pixels and a short edge of 4320 pixels; very large sources or enlargement factors may exceed those limits.
  • Solution: Reduce the enlargement factor or resize the source earlier in your workflow. Do not expect an automatic bucket to downscale an already oversized source.

NVIDIA runtime DLLs are missing or cannot be loaded

  • Explanation: This is a descriptive setup issue; the exact loader message is not included. The extension does not redistribute NVIDIA runtime DLLs, and processing needs compatible files in the resolved runtime folder. Enlargement beyond 1× also requires the standard DLSS runtime.
  • Solution: Supply the required compatible DLLs and set runtime_dir to their containing folder, or configure DLSS5_RUNTIME_DIR. If native-size enhancement works but enlargement fails, check that nvngx_dlss.dll is present as well.

Wine prefix validation or Linux backend startup fails

  • Explanation: This is a descriptive setup issue. Linux processing requires a usable Wine environment, and the source specifies DXVK and DXVK-NVAPI as required prefix components. A path to an unconfigured or different prefix can prevent startup.
  • Solution: Verify that wine_prefix points to the intended configured prefix and that its required components are installed. If the field is empty, check DLSS5_WINEPREFIX and then ~/.wine to confirm which environment is being used.

Selected backend cannot run in the current environment

  • Explanation: This is a descriptive compatibility issue. The Linux backend depends on Wine, while the Windows bridge uses an in-process D3D12 path; manually selecting a backend does not install its dependencies or make it portable to another environment.
  • Solution: Return backend to auto. If you intentionally force a backend, first verify that its operating-system requirements and supporting runtime components are available.

The enhancement host did not send END1

  • Explanation: This exact Linux-backend error means the worker did not send the expected completion marker when processing ended. The message identifies an incomplete or unexpected host shutdown, but does not by itself establish the underlying cause.
  • Solution: Inspect the preceding console messages for the original failure. Verify the runtime folder and Wine setup, then retry with one small image to separate environment problems from workload-related failures.

The enhancement host failed (exit N)

  • Explanation: The source generates The enhancement host failed (exit `{rc}`) with the actual nonzero exit code substituted. This means the worker exited unsuccessfully; the exit number alone is not enough to identify the reason.
  • Solution: Check earlier worker output, verify DLL compatibility and Linux dependencies, and reproduce the problem with a small input. Include the actual exit code and preceding log messages when seeking support.

chunked decode unavailable (...); full decode

  • Explanation: This console-message pattern comes from the provided source. It is a fallback warning rather than necessarily a processing failure: the node could not create its chunked reader and will attempt full decoding instead. Large clips may then require much more memory.
  • Solution: Read the exception inside the parentheses for the reader problem. Test a shorter or smaller clip, verify the source file can be decoded in your environment, and check available memory before retrying a long render.

Red and blue appear swapped

  • Explanation: This is a visual symptom, not a quoted exception. Automatic channel-layout detection may not choose the correct interpretation for the output, causing obvious red/blue reversal.
  • Solution: Change channel_order from auto to RGBA or BGRA and compare a frame containing recognizable colors. Use the setting that restores the expected colors.

Unrelated images influence one another, or temporal artifacts appear

  • Explanation: This is a visual troubleshooting category. temporal sequence deliberately retains history between frames, which is inappropriate for unrelated pictures. In footage, unsuitable motion estimation or missed scene resets can also affect temporal behavior.
  • Solution: Use still images for independent artwork. For coherent footage, test motion = auto or optical_flow and, where applicable, lower scene_change_threshold to trigger resets more readily at cuts.

Changing skin, global_tone, or intensity appears to do nothing

  • Explanation: These controls have documented limitations rather than necessarily indicating a failure. skin is inactive with auto_mask = off; global_tone has no effect on the tested 310.8 runtime; and intensity values above 1.0 usually provide no extra effect.
  • Solution: Enable auto_mask when testing skin adjustments, do not rely on global_tone with the documented test runtime, and compare intensity values below 1.0 when checking its influence. Use detail carefully if you specifically want stronger compositing of the model’s changes.

Upscaling is ignored, or audio remains despite keep_audio = off

  • Explanation: These can be expected bypass behaviors. style = off (bypass NR) disables all processing, including upscaling, and a file-backed VIDEO is returned unchanged, including its existing audio.
  • Solution: Select an active rendering style to apply enlargement and rebuild the video with the requested audio setting. If you want unchanged pictures but no audio, remove the audio in a separate video-processing step rather than relying on file-backed bypass.

RH DLSS5 Enhance (NR / Upscale) Related Nodes

Go back to the extension to check out more related nodes.
ComfyUI-RH-DLSS5
RunComfy
Copyright 2025 RunComfy. All Rights Reserved.

RunComfy is the premier ComfyUI platform, offering ComfyUI online environment and services, along with ComfyUI workflows featuring stunning visuals. RunComfy also provides AI Models, enabling artists to harness the latest AI tools to create incredible art.

RH DLSS5 Enhance (NR / Upscale)