DLSS 5 is NVIDIA's neural rendering pass. In a game, it reconstructs the material and lighting detail that real-time rendering has to leave out, such as skin subsurface scattering, light transmission through hair, and contact shadows. This node pack drives that same native renderer over video frames and image batches, with optional 1x to 16x RTX upscaling, DLSS Ray Reconstruction, and DLSS Frame Generation.

It is not a filter that imitates a look. Frames go to the native renderer and come back reconstructed. The model reconstructs materials and recovers structure (skin, hair, fabric) that the source only implied, rather than just sharpening edges.

Unlike older community implementations that relied on ReShade injection, this pack drives the official v17.5 headless CLI (VE_CLI.exe), providing a stable, native, and highly optimized pipeline.

Acknowledgment: This node pack acts as a bridge to the excellent DLSS 5 Visual Enhancer project by Merserk. All underlying AI rendering, upscaling, and frame generation capabilities are powered by their work and the official NVIDIA runtime. This project simply makes that powerful technology accessible directly within ComfyUI workflows.

🚀 What's New in v1.3.1

This update focuses on robust video handling, seamless integration with VideoHelperSuite, and turning the Video Path Selector into a complete, all-in-one video loader.

New & Improved Features

Hybrid Video Processing (Images + Audio): The DLSS5 Enhance (Video) node now accepts optional Images (IMAGE) and Audio (AUDIO) inputs. It encodes the image batch into a temporary video for flicker-free DLSS processing, seamlessly muxes your audio onto it, processes the video through the CLI, and finally attaches the audio to the enhanced output.
All-in-One Video Path Selector: The Selector node now outputs images and audio alongside video_path. It acts as a complete video loader, allowing you to bypass external loader nodes for simple workflows.
Robust Frame Loading: Switched the Selector's internal frame reader to a robust ffmpeg subprocess method. This completely fixes "Could not load meta information" errors on certain MP4 codecs that imageio struggled with.
Memory Safety Widget: Added a frame_load_cap widget to the Selector to prevent RAM crashes when loading long videos.
VideoHelperSuite Path Compatibility: Fixed path resolution to support VideoHelperSuite's parent-relative paths (e.g., ComfyUI/input/video/...), allowing seamless chaining between the Selector and VHS nodes.
Final Preview Node: Added OUTPUT_NODE = True to the Selector, preventing the "Prompt has no outputs" warning when it is used as the final node in a chain.
UI & Safety Improvements: Moved Video Path to the top of the Enhance (Video) node for better workflow visibility. Increased the default preview height in the frontend. Added a safety check to prevent the DLSS CLI from crashing when fed a 1-frame batch.

🧠 How It Works

This pack acts as a bridge between ComfyUI and the native DLSS 5 Visual Enhancer CLI.

Preset Chain: The control nodes (Pro Controls, Color Enhancement, Sharpening & Grain) build and pass along a JSON preset string (pro_controls). Each node injects its settings and enables the required pipeline stages (neural_model, coloring, cas_sharpening, grain, dlss_super_resolution, etc.).
Neural Rendering & Ray Reconstruction: The Enhance node exports your frames (or passes your video path) to VE_CLI.exe with the preset via --preset. The CLI hosts the native NVIDIA runtime to reconstruct high-frequency material details. With Ray Reconstruction enabled, the dedicated reconstruct command runs at native resolution (1x) with high-precision color handling.
RTX Upscaling: Reconstructed frames optionally pass through NVIDIA's RTX Super Resolution engine (up to 16x).
DLSS Frame Generation: Optionally interpolates video to a target FPS (RTX 40/50 only).
Color Match (Post): The Color Matcher node compares the enhanced output against a reference image and applies a Reinhard color transfer to cancel any color drift introduced by the neural pass.

Requirements

| Item | Requirement | | --- | --- | | GPU | NVIDIA RTX 20, 30, 40, or 50 Series. (GTX, AMD, and Intel GPUs are not supported). | | GPU (Frame Gen) | DLSS Frame Generation requires RTX 40 or 50 Series + Hardware-accelerated GPU scheduling. | | Driver | Recent NVIDIA Game Ready or Studio Driver (560.xx+). | | OS | Windows only. | | ComfyUI | Recent build supporting standard IMAGE and AUDIO tensor types. | | Python | imageio-ffmpeg, scikit-image (installed via requirements.txt). |

️ Installation

Step 1: Install the Node Pack Clone this repository into your ComfyUI/custom_nodes/ folder:

git clone https://github.com/Strix-3D/ComfyUI-DLSS5-Enhancer-v15.git

Then install dependencies from your main ComfyUI folder:

.\python_embeded\python.exe -m pip install -r ComfyUI\custom_nodes\ComfyUI-DLSS5-Enhancer-v15\requirements.txt

Step 2: Download the Runtime & Accept the License

.\python_embeded\python.exe ComfyUI\custom_nodes\ComfyUI-DLSS5-Enhancer-v15\install.py

Wait for the ~700MB runtime to download into runtime/.

Step 3: Restart ComfyUI. Nodes appear under the DLSS5_v17 category.

Nodes & Parameters

Nodes Overview

1. DLSS5 Enhance (Image)

Applies Neural Rendering (or Ray Reconstruction) and optional RTX upscaling to an image batch.

| Widget | Options / Range | Default | Effect | | --- | --- | --- | --- | | Images | IMAGE | — | Input image or batch. | | NR Style | Default / Natural / Cinematic | Default | The AI's reconstruction "personality". Natural is best for faces; Cinematic adds punch but can crush shadows. | | NR Intensity | 0.0 – 2.0 | 1.0 | Strength of the neural effect. >1.5 risks AI hallucinations. | | NR Passes | 1 – 4 | 1 | Neural render passes. More passes = more reconstructed micro-detail. | | Scale | 1.0 – 16.0 | 2.0 | RTX upscaling multiplier. 1.0 skips upscaling. | | Ray Reconstruction | BOOLEAN | False | Native 1x reconstruction with high-precision color. For 3D/RTX captures. | | Pro Controls (optional socket) | STRING | — | Input from the control chain. |

Output: IMAGE

2. DLSS5 Enhance (Video)

Processes a video file natively in a single CLI pass (temporally stable), with optional upscale, interpolation, and audio muxing. Now supports hybrid Image+Audio input.

| Widget | Options / Range | Default | Effect | | --- | --- | --- | --- | | Video Path (Top) | STRING (linkable) | — | Relative path in input/, or linked from Selector/other nodes. (Functions optionally if Images are provided). | | NR Style / NR Intensity / NR Passes | as above | Default / 1.0 / 1 | Neural render personality, strength, passes. | | Scale | 1.0 – 16.0 | 1.0 | RTX upscale multiplier. | | Codec | HEVC / H264 / AV1 / ProRes Proxy | H264 | Hardware encoder. H264 caps at 4096px wide. | | Container | MKV / MP4 / MOV | MP4 | Output container. | | Encoder Quality | Auto / Good / Best / Max | Auto | Encoding bitrate/quality. | | Target FPS | 0 – 120 | 0 | 0 = disabled. DLSS Frame Generation target (RTX 40/50). | | Ray Reconstruction | BOOLEAN | False | Mutually exclusive with DLSS Super Resolution. | | Optical Flow Quality | Disabled / High / Medium / Low | Disabled | Temporal stability for Ray Reconstruction. | | Filename Prefix | STRING | DLSS5 | Output filename prefix (timestamp appended). | | Output Directory | STRING | — | Relative subfolder inside output/. | | Copy Audio | BOOLEAN | True | Mux original audio into the output. | | Input FPS | FLOAT | 24.0 | Frame rate used ONLY when encoding an IMAGE batch into a temporary video. | | Images (optional socket) | IMAGE | — | Pass an IMAGE batch (e.g. from VideoHelperSuite). The node will encode it to a temp video for flicker-free processing. | | Audio (optional socket) | AUDIO | — | Audio track to mux into the temporary video when using an IMAGE batch. | | Pro Controls (optional socket) | STRING | — | Input from the control chain. |

Output: output_path (STRING)

3. DLSS5 Video Path Selector

Browse, upload, and chain video paths without typing. Now acts as a complete video loader with live preview, content deduplication, and robust frame extraction.

| Widget / Output | Type | Effect | | --- | --- | --- | | Video Path | STRING (linkable) | Type a relative path, upload, or accept a link. | | frame_load_cap | INT | 0 = load all frames. Limit this to prevent memory crashes on long videos. | | Choose file to upload... | Button | Uploads into input/; reuses identical existing files by SHA-1 hash. | | Live preview | Preview | Updates on upload, on typing, and on execution. | | Output: video_path | STRING | The resolved relative path. | | Output: images | IMAGE | The loaded video frames (batch). | | Output: audio | AUDIO | The extracted audio track. |

4. DLSS5 Video Info

Lightweight metadata probe for chaining. Plug in any path (e.g., an Enhance node's output_path). Outputs: video_info (JSON), width, height, fps, frame_count.

5. DLSS5 Pro Controls

Advanced native render controls. Starts the pro_controls chain.

| Widget | Range / Options | Default | Effect | | --- | --- | --- | --- | | Local Tone Strength | 0.0 – 1.0 | 1.0 | Local tone intensity. | | Local Structure Strength | 0.0 – 1.0 | 1.0 | Local structure intensity. | | Automatic Mask | BOOLEAN | True | Targets skin enhancement to detected skin regions. | | Skin Structure Strength | 0.0 – 1.0 | 1.0 | Skin detail (active when Automatic Mask is ON). | | DLSS Super Resolution | BOOLEAN | False | Enables the in-render DLSS SR stage. Keep Enhance Scale at 1.0 when using it. | | DLSS Mode | 1x (DLAA) / 1.5x / 1.724x / 2x / 3x | 1x (DLAA / native) | DLSS render scale (only when SR is ON). | | DLSS Preset | Default / J / K / L / M | M | DLSS model preset (only when SR is ON). |

Output: pro_controls (STRING)

6. DLSS5 Color Enhancement

Full Light & Color grading panel. Injects lut_* values and enables the coloring stage.

| Widget | Range | Default | Widget | Range | Default | | --- | --- | --- | --- | --- | --- | | Exposure | -3.0 – +3.0 | 0.0 | Temperature | -100 – 100 | 0 | | Contrast | -100 – 100 | 0 | Tint | -100 – 100 | 0 | | Highlights | -100 – 100 | 0 | Hue Shift | -180 – 180 | 0 | | Shadows | -100 – 100 | 0 | Vibrance | -100 – 100 | 0 | | Whites | -100 – 100 | 0 | Saturation | -100 – 100 | 0 | | Blacks | -100 – 100 | 0 | | | | | Midtones | -100 – 100 | 0 | | | |

Input: Pro Controls (optional) • Output: pro_controls (STRING)

7. DLSS5 Sharpening & Grain

Toggleable sharpening and film grain. Toggles add/remove the cas_sharpening and grain pipeline stages.

| Widget | Range / Options | Default | Effect | | --- | --- | --- | --- | | Sharpening | BOOLEAN | True | Enable the sharpening stage. | | Sharpening Method | NVIDIA NIS / AMD CAS | NVIDIA NIS | NIS recovers edges from soft sources; CAS boosts micro-contrast. | | Sharpness | 0 – 100 | 50 | Sharpening intensity. | | Grain | BOOLEAN | True | Enable the grain stage. | | Grain Amount | 0 – 100 | 20 | Grain intensity. | | Grain Size | 0.1 – 5.0 | 1.0 | Grain particle size. | | Color Grain | 0 – 100 | 0 | Monochrome (0) to full-color grain. | | Tonal Response | 0 – 100 | 70 | How grain reacts to luminance. | | Grain Seed | 0 – 999999 | 0 | Random seed for reproducibility. | | Grain Animated | BOOLEAN | True | Animates grain over time (video). |

Input: Pro Controls (optional) • Output: pro_controls (STRING)

8. DLSS5 Color Matcher

Place AFTER the Enhance node. Cancels neural-render color drift so the final output matches your reference's colors. Pure image processing (Reinhard transfer in Lab space via scikit-image) — no CLI/LUT involved. Resolution-independent (statistics-based), so the reference can be any size. When Source and Reference are the same image, the result is an exact identity (no change).

| Widget | Type | Effect | | --- | --- | --- | | Source Image | IMAGE | The enhanced/drifted image (Enhance node output). | | Reference Image | IMAGE | The colors you want (usually the original input). | | Match Strength | 0.0 – 1.0 | Blend between uncorrected (0) and fully matched (1). |

Output: matched_image (IMAGE)

🔗 Recommended Control Chain

DLSS5 Pro Controls ➔ DLSS5 Color Enhancement ➔ DLSS5 Sharpening & Grain ➔ Enhance (Image/Video)
                                                                                        │
Load Image  Enhance (Images)                    Enhance (IMAGE out) ➔ Color Matcher (Source)
Load Image ────────────────────────────────────────────────────────➔ Color Matcher (Reference)

The chain is optional — the Enhance nodes work standalone. Muted nodes are skipped entirely; a muted control node contributes nothing (its settings are silently ignored).

⚙️ CLI Flag Mapping (Advanced)

| ComfyUI Widget | CLI Command / Flag | | --- | --- | | NR Style | --style 0 / 1 / 2 | | NR Intensity | --strength [0.0–2.0] | | NR Passes | preset key nr_passes | | Ray Reconstruction | uses reconstruct command instead of render | | Optical Flow Quality | --flow-quality high/medium/low | | Scale | --scale [1.0–16.0] | | Target FPS | interpolate command with --fps | | Codec / Container / Encoder Quality | --codec / --container / --quality | | Control chain settings | --preset (JSON, schema auto-normalized to 38) |

Hardware Capability Cheat Sheet

| GPU Series | Neural Rendering | RTX Upscaling (16x) | DLSS Frame Generation | | --- | --- | --- | --- | | RTX 50 Series | ✅ | ✅ | ✅ | | RTX 40 Series | ✅ | ✅ | ✅ | | RTX 30 Series | ✅ | ✅ | ❌ | | RTX 20 Series | ✅ | ✅ | ❌ | | GTX / AMD / Intel | | ❌ | ❌ |

✅ Recommended Settings

| Goal | Settings | | --- | --- | | Clean up generated video, same resolution | Scale 1.0, NR Style Default, NR Intensity 1.0 | | Clean up and upscale | Same, with Scale 2.0+ | | Smooth 30fps → 60fps | Target FPS 60 (RTX 40/50 only) | | Most realistic faces | NR Style Natural, NR Intensity 0.8, Pro Controls Auto Mask ON | | Maximum detail reconstruction | NR Passes 3–4 (watch for hallucinations) | | High-precision 3D/RTX cleanup | Ray Reconstruction ON, NR Style Default | | More contrast/punch | NR Style Cinematic (check shadows) | | Cancel AI color drift | Color Matcher after Enhance; Reference = original input | | Output too strong/plastic | Drop NR Intensity to 0.6–0.7 or switch to Natural |

🧪 Example Workflows

Image Enhancement + Color Match

Image Workflow

Load Image ➔ Pro Controls ➔ Color Enhancement ➔ Sharpening & Grain  Enhance (Image)
                                                                              │
Load Image ➔ Enhance (Images)                                                 │
Load Image ➔ Color Matcher (Reference)          Enhance out  Color Matcher (Source) ➔ Save/Compare

Video File Enhancement (Path based)

Video Workflow (Path)

Video Path Selector ➔ Enhance (Video) ➔ output_path ➔ Video Info / Load Video (Path) ➔ Video Combine

Video File Enhancement (Hybrid Image/Audio based)

Video Workflow (Hybrid)

Load Video (Path) [IMAGE out] ──┐
Load Video (Path) [audio out] ──┼──➔ Enhance (Video) [Images & Audio inputs] ➔ output_path

Troubleshooting

| Message / Issue | Cause and Fix | | --- | --- | | DLSS CLI requires at least 2 frames... | Caused by a failed video loader sending a dummy 1-frame image. Use VideoHelperSuite's Load Video (Path) or check the Selector's frame_load_cap widget. | | Open host video encoder: Invalid argument | Usually caused by odd output dimensions or feeding the CLI a 1-frame file. The node now auto-pads odd dimensions and checks frame counts. | | Unsupported preset schema version | Fixed in v1.3.0 — presets are normalized to schema 38 automatically. If it reappears, update the pack. | | Upload creates duplicate copies | Fixed in v1.3.0 — uploads are deduplicated by SHA-1 content hash; existing files in input/ are reused. | | Control settings seem ignored | Check the node isn't muted (faded). Muted nodes are skipped and contribute nothing. | | Color Matcher output = source | Expected when Source and Reference are the same image (identity transfer). | | ModuleNotFoundError: skimage | Run pip install -r requirements.txt in your ComfyUI environment. | | SmartScreen blocks VE_CLI.exe | Right-click the .exe → Properties → Unblock. | | Width N exceeds 4096 | H264 encoder cap. Switch Codec to HEVC/AV1 or lower Scale. | | Frame Generation fails on RTX 20/30 | Requires RTX 40/50. Set Target FPS to 0. | | Video preview stale on linked path | Preview refreshes on execution; hard-refresh browser (Ctrl+F5) if needed. | | Output looks "plastic" | NR Intensity too high. Drop to 0.7–1.0 or use Natural style. |

Project Layout

ComfyUI-DLSS5-Enhancer-v15/
   install.py             Downloads the runtime and prompts for license acceptance
   nodes.py               ComfyUI node definitions (8 nodes)
   paths.py               Runtime discovery and validation
   __init__.py            ComfyUI extension entrypoint (serves web/)
   README.md              This file
   LICENSE                MIT License (Python source)
   requirements.txt       Python dependencies (imageio-ffmpeg, scikit-image)
   images/                Workflow screenshots for documentation
   web/dlss5_selector.js  Frontend extension (upload button + live preview)
   runtime/               Native runtime (VE_CLI.exe, NVIDIA DLLs) - *Not tracked*
   config.json            Local path configuration - *Not tracked*

️ Licensing and Attribution

The Python source code in this repository is licensed under the MIT License.

This repository does not host or redistribute the AI runtime binaries. The install.py script downloads the official DLSS 5 Visual Enhancer v17.5 release from Merserk/dlss5-visual-enhancer. The downloaded runtime contains proprietary components owned by their respective creators:

NVIDIA Corporation: DLSS Super Resolution, Neural Rendering (v310.9.16.0), Frame Generation, Ray Reconstruction, Optical Flow, and RTX Video SDKs (Proprietary).
Merserk: Visual Enhancer CLI and Engine (Merserk Source License).
FFmpeg/mpv: Media processing binaries (GPLv2/v3).

By installing this node pack, you agree to use these components only under the terms permitted by their respective licenses. This project is not affiliated with or endorsed by NVIDIA or the Merserk project.