DLSS 5 Visual Enhancer
Windows application for applying a DLSS 5 Neural Rendering feature-18 pipeline to images and video, NVIDIA DLSS Frame Generation to video frame interpolation, and DLSS 5 Neural Rendering during Live video playback through a local Gradio interface. It is an independent community project and is not affiliated with, sponsored by, or endorsed by NVIDIA, ReShade, RenoDX, FFmpeg, or their contributors.
Original
https://github.com/user-attachments/assets/8df8bd4c-01b4-47dd-9705-3614a0b0ff75
DLSS 5
https://github.com/user-attachments/assets/cff68783-4ee9-4c99-8b36-4eee2a6437ec
Frame Generation (DLSSG)
https://github.com/user-attachments/assets/81c29005-e4f0-4acf-b9f7-d58850bb055f
Installation
- Download the latest release.
- Unpack the downloaded ZIP archive.
- Run
start.bat.
Main features
- Images: single-image and batch processing with per-file success/failure results, responsive input/output previews, individual downloads, a ZIP of successful outputs, batch manifests, and diagnostic JSON reports.
- Image formats: common Pillow formats plus HEIF/HEIC, SVG, and many camera RAW formats. Outputs are PNG, JPEG, WebP, AVIF, or TIFF.
- Image handling: EXIF orientation is applied; ICC input is converted to sRGB; supported EXIF, DPI, and XMP metadata is retained; alpha is preserved except when JPEG composites it over white. EXIF/TIFF rational values and other unusual metadata are converted safely for diagnostic JSON without modifying the metadata used to encode the output. Animated and multipage files use only the first frame/page.
- Video: single-video and batch-video Neural Rendering, plus one-frame and three-second previews for a single selected video. H.264, H.265, AV1, and ProRes Proxy are available in MP4, MKV, or MOV where compatible, each in a plain CPU variant or an
(NVIDIA NVENC) GPU variant (except ProRes Proxy, which is CPU-only).
- Frame Interpolation: single-video and batch-video processing with NVIDIA DLSS Frame Generation, selectable output rates from 23.976 to 480 FPS, and a three-second preview for a single selected video. The processing pipeline now overlaps more of the decode, motion-guide, Frame Generation, and encode work for substantially higher throughput.
- Live: watch local videos, direct network streams, YouTube, and Twitch with DLSS 5 Neural Rendering applied during playback. NR Preset, NR Style, intensity, tone, structure, skin structure, Automatic Mask, and DLSS Model Preset changes can be applied while playback is running without restarting the Live session.
- Live playback controls: Source quality, processing height from 480p to 4K, Auto/Source/60/30/24 FPS modes, adjustable playback buffering, Fast or Quality motion guides, and optional automatic playback in the bundled MPV player. Upscaling and source/playback configuration changes apply on the next Start.
- Direct disk batch processing: Image, Video, and Frame Interpolation tabs accept an absolute Input path for a file or folder and an optional Output path. Folder processing uses supported files in that folder only, sorted by filename. When either path is used, previews and browser downloads are disabled and results are written directly to disk.
- Batch progress: every file has its own Queued, Running, Complete, Failed, Cancelled, or Skipped state with percentage, elapsed time, output path, and processing details. Completed files are preserved if a batch is stopped, while the active incomplete output is cleaned up safely.
- Single-only video previews: input and output players are shown when exactly one uploaded video is selected. They are hidden for multi-video batches and direct disk mode. The final-render player follows the Preview Encoding setting: Auto may show an H.264 derivative when the result itself is not browser-playable, while Disabled always shows the actual file.
- Media preservation: frame timestamps and display rotation are handled; original metadata and chapters are copied. MKV copies compatible audio and subtitle streams, while MP4/MOV convert audio to 192 kbps AAC. Frame Interpolation also preserves supported text subtitles in MP4/MOV.
- GPU selection: AI Processing and Video Processing GPUs can be selected separately. The AI GPU is used for DLSS Neural Rendering and Frame Generation; the Video GPU is used only for codecs suffixed
(NVIDIA NVENC), while plain H.264/H.265/AV1 and ProRes Proxy remain CPU-based. Automatic selection is available for both.
- DLSS 5 Architecture: selectable Neural Rendering runtime builds are available for Turing and higher, Ada Lovelace and higher, and Blackwell and higher. Auto selects the matching build from the chosen AI Processing GPU, with RTX 20/30 using Turing+, RTX 40 using Ada Lovelace+, and RTX 50 using Blackwell+.
- Preview Encoding: a Settings-tab control for how in-app video previews are produced on the Video and Frame Interpolation tabs (preview buttons and final-render player). Auto uses the result directly when the browser can play it (MP4 + H.264, verified by probe) and otherwise creates an H.264 preview; Always H.264 always generates a browser-compatible preview; Disabled never creates one and sends the actual file to the browser. Non-H.264 previews can be slower and larger.
- HDR Mode: an opt-in 10-bit output that copies the input colorspace and keeps HDR when the input is HDR. It is available only for H.265, AV1, and ProRes Proxy; H.264 stays 8-bit SDR. Frame Interpolation accepts SDR video only unless HDR Mode is enabled.
- GPU compatibility: RTX compatibility is no longer restricted by a fixed Ampere, Ada, and Blackwell allowlist. RTX hardware is detected first, while architecture information is used for runtime selection and diagnostics. Actual DLSS and Frame Generation feature availability is determined by the installed NVIDIA runtime and driver.
- Render metadata: Image and Video Neural Rendering outputs can store the applied DLSS 5 settings in supported output metadata without replacing existing user descriptions. Video metadata notes are supported for MP4 and MKV outputs.
- Renaming: Image, Video, and Frame Interpolation outputs support Auto timestamped naming, Copy naming with the original base filename, or a Custom suffix. Existing outputs are never overwritten.
- Safety and diagnostics: only one GPU render runs at a time. Stop cancels active workers and encoders, completed batch outputs are retained, incomplete outputs are cleaned up, and finished files are published only after the relevant render path and output properties are verified.
- Persistent controls: Image, Video, and Live share Neural Rendering settings and the DLSS Model Preset. Frame Interpolation settings, HDR Mode selections, Preview Encoding, DLSS 5 Architecture, and GPU selections are also saved in
config.ini. Settings presets can be exported to or imported from JSON, and Reset restores the relevant controls to their defaults. Live source, quality, buffering, and playback controls are session-local.
The application creates outputs/, logs/, jobs/, and live/ when needed. Successful media is written to outputs/, reports and manifests to logs/, temporary active-render data to jobs/, and temporary Live session data to live/. Old cache and abandoned Live session data are cleaned up automatically.
Requirements
- 64-bit Windows 11 with Direct3D 12.
- NVIDIA GeForce RTX GPU with a compatible NVIDIA driver. Auto DLSS Architecture selection recognizes RTX 20/30, RTX 40, and RTX 50 generations, while unusual or newer RTX hardware is allowed to continue to runtime feature detection instead of being rejected only because its architecture metadata is unknown.
- Frame Interpolation requires a GPU, driver, and NVIDIA DLSS Frame Generation runtime combination that reports Frame Generation support. Hardware-accelerated GPU scheduling (HAGS) should be enabled; if it is disabled, the app reports it as a diagnostic and lets the runtime determine whether Frame Generation can start.
- Live local files and direct stream URLs use the portable media tools included with the application. YouTube and Twitch page URLs require the bundled
yt-dlp, while automatic playback requires the bundled MPV player or can be disabled with Open in MPV.
Settings
| Neural control |
Values |
Default |
| NR Preset |
Default, Preset #1, Preset #2, Preset #3 |
Default |
| NR Style |
Default, Natural, Cinematic |
Default |
| NR Intensity |
0.00–2.00 |
1.00 |
| Local Tone Strength |
0.00–2.00 |
1.00 |
| Local Structure Strength |
0.00–2.00 |
1.00 |
| Skin Structure Strength |
-1.00–2.00 |
-1.00 (native default) |
| Automatic Mask |
Off, On |
Off |
The preset numbers are experimental native model hints; their visual effect is content-dependent.
| DLSS Model Preset |
SDK value |
Behavior |
| Default |
0 |
Applies no override and lets NVIDIA choose its normal mode-specific preset |
| J |
10 |
Forces preset J for every supported DLSS scaling mode |
| K |
11 |
Forces preset K for every supported DLSS scaling mode |
| L |
12 |
Forces preset L for every supported DLSS scaling mode |
| M |
13 |
Forces preset M for every supported DLSS scaling mode |
The DLSS Model Preset is independent of the experimental Neural Rendering NR Preset control.
| Upscaling mode |
Factor |
Behavior |
| DLAA / native |
1× |
Keeps the source dimensions |
| Quality |
1.5× |
Produces 1.5× output dimensions |
| Balanced |
1.724× |
Produces approximately 1.724× output dimensions |
| Performance |
2× |
Produces 2× output dimensions |
| Ultra Performance |
3× |
Produces 3× output dimensions |
Output dimensions are rounded to even pixels and limited to a 7680×4320 boundary.
| Frame Interpolation setting |
Choices and behavior |
Default |
| Output FPS |
23.976, 25, 29.97, 30, 50, 59.94, 60, 90, 119.88, 120, 144, 165, 180, 240, 360, or 480 FPS |
60 |
| DLSS engine |
Auto, Native DLSSG, or Cascade |
Auto |
| Video codec |
H.264, H.265, AV1, or ProRes Proxy, each in a plain CPU variant or an (NVIDIA NVENC) GPU variant (ProRes Proxy is CPU-only) |
H.264 |
| Container |
MP4, MKV, or MOV; ProRes Proxy requires MKV or MOV |
MP4 |
| Encoding quality |
Auto (Default), Good, Best, or Max |
Auto (Default) |
| HDR Mode |
Off, On; 10-bit output that copies the input colorspace; H.265 / AV1 / ProRes only |
Off |
| Rename |
Auto adds a DLSSFG timestamp; Copy keeps the original base name; Custom appends the entered suffix |
Auto |
Auto uses a supported exact native DLSSG grid when possible and the cascade path when required. If the selected output FPS is equal to or below the source rate, source frames are resampled without generating additional frames.
| Live setting |
Choices and behavior |
Default |
| Source |
Online URL or Local video |
Online |
| Source quality |
Auto, 480p, 720p, 1080p, 1440p, or 2160p |
Auto |
| Max input height |
480p, 720p, 1080p, 1440p, or 2160p before DLSS processing |
720p |
| Segment length |
1, 2, or 4 seconds |
2 seconds |
| Live frame rate |
Auto, Source, 60, 30, or 24 FPS |
Auto |
| Playback buffer |
2–30 seconds |
6 seconds |
| Motion guide quality |
Fast or Quality |
Fast |
| Open in MPV |
Off, On |
On |
Live Neural Rendering controls and the DLSS Model Preset are shared with the Image and Video tabs. Effect controls can update during playback, while the upscaling factor and Live source/playback controls take effect on the next Start.
| Output setting |
Choices and behavior |
| Image format |
PNG/TIFF are lossless; JPEG/WebP/AVIF use the 1–100 quality control (default 95) |
| Video codec |
H.264, H.265, AV1, or ProRes Proxy, each in a plain CPU variant or an (NVIDIA NVENC) GPU variant (ProRes Proxy is CPU-only) |
| Container |
MP4, MKV, or MOV; ProRes Proxy requires MKV or MOV |
| Encoding quality |
Auto (Default) uses resolution/FPS/codec; Good = Auto×2; Best = Auto×4; Max uses CQ/CRF 0; ProRes uses its fixed Proxy profile |
| HDR Mode |
Off, On; 10-bit output that copies the input colorspace; H.265 / AV1 / ProRes only |
| Rename |
Auto adds the DLSS5 timestamp; Copy keeps the original base name; Custom appends the entered suffix before the output extension |
Plain H.264/H.265/AV1 are CPU-based (libx264 / libx265 / libsvtav1) and never require a GPU. Only codecs suffixed (NVIDIA NVENC) need working NVENC support on the selected or automatically chosen Video Processing GPU at the requested output size. ProRes Proxy uses CPU-based 10-bit 4:2:2 encoding, although the verified Neural Rendering path remains RGBA8.
| Application setting |
Choices and behavior |
| AI Processing GPU |
Automatic or a compatible NVIDIA RTX GPU; used for Neural Rendering and Frame Generation |
| Video Processing GPU |
Automatic or an available NVIDIA GPU; used only for codecs suffixed (NVIDIA NVENC), while plain H.264/H.265/AV1 and ProRes stay on CPU |
| Preview Encoding |
Auto (default), Always H.264, or Disabled; controls how in-app video previews are produced on the Video and Frame Interpolation tabs |
| DLSS 5 Architecture |
Auto, Turing and higher, Ada Lovelace and higher, or Blackwell and higher; Auto selects the runtime build from the AI Processing GPU |
| Settings preset |
Export all adjustable Image, Video, Frame Interpolation, GPU, HDR Mode, Preview Encoding, and DLSS 5 Architecture settings to JSON, or import a preset to apply and save them |
Saved GPU selections use stable GPU identity. If a previously saved GPU is no longer available, that selection returns to Automatic rather than silently switching to another saved device.
License and third-party notices
Original application code is licensed under the MIT License, copyright © 2026 Merserk. That license covers only original project code; it does not relicense or grant rights to any third-party software, model, binary, trademark, media, or other asset.
- NVIDIA DLSS/NGX: NVIDIA and its suppliers retain their rights in genuine NVIDIA SDK and runtime files, including DLSS Neural Rendering and DLSS Frame Generation components. Use and distribution are governed by the NVIDIA RTX SDK License. Their presence in a portable package does not imply a standalone redistribution right, and this project must not be represented as NVIDIA-sponsored or endorsed.
- FFmpeg: the referenced Gyan.dev
9.0.1-full_build was configured with GPL and version-3 components and is distributed under GPLv3. Its build information, license, and exact corresponding FFmpeg source commit are retained under bin/ffmpeg/. Anyone redistributing that binary must satisfy the applicable GPLv3 and corresponding-source obligations. See FFmpeg licensing.
- ReShade: copyright belongs to Patrick Mours and contributors; ReShade is available under the BSD 3-Clause License.
- RenoDX: RenoDX core is copyright its authors and available under MIT. This does not establish the license of the separate
renodx-dlss5.addon64 file.
- MPV and yt-dlp: Live mode can use the bundled portable MPV player and yt-dlp resolver. They and their dependencies retain their own copyright and license terms; preserve the notices shipped with each distribution.
- Python and packages: Python is provided under the PSF License. Gradio, Pillow, pillow-heif, rawpy, resvg-py, PyAV, OpenCV, NumPy, their transitive dependencies, and bundled codecs retain their own copyright and license terms; preserve the notices shipped with each distribution.
NVIDIA, GeForce RTX, NGX, and DLSS are trademarks and/or registered trademarks of NVIDIA Corporation. FFmpeg, ReShade, RenoDX, MPV, yt-dlp, Python, and other names belong to their respective owners. Codec patent or other permissions may also be required depending on jurisdiction and use. Review the controlling licenses before building or distributing a complete package.
DLSS 5 Visual Enhancer 是一款 Windows 应用，可将 NVIDIA 的 DLSS 5 Neural Rendering 应用于图片和视频，用 DLSS Frame Generation 做视频插帧，并在视频播放期间实时应用 Neural Rendering——这一切都通过本地网页界面完成。