A free, open-source Windows player that runs your videos, photos and GIFs through NVIDIA's DLSS 5 neural renderer on your own RTX card while you watch, keeps the original one key away, on the same frame, and saves the result as a file.
Windows · NVIDIA RTX · driver 610.47+ · MIT licensed. A community project, not an NVIDIA product. Is it safe?
Drag the seam above. Both sides are one frame of the 007 First Light story trailer: the source on the left, the player's render at its default settings on the right, cropped identically and not retouched, sharpened or colour-graded. The change sits in skin, stubble and light across the face: the renderer rebuilds detail rather than laying a filter over the picture.
Source2560×1440 · 59.94 fps · 96.5 s
Renderthe whole video, at the default settings, upscaling off
MachineRTX 4080 SUPER · driver 610.47
007 First Light and Grand Theft Auto VI at the default settings: a paused frame split down the face, GTA VI playing with the divider held still, then the player's Difference, Side by side and loupe views, and a live render filling the timeline (shown at 2× speed). No sound.
Rendered on an RTX 4080 SUPER; how it was made. 007 First Light — IO Interactive; Grand Theft Auto VI — Rockstar Games.
Each still is one frame of a trailer, the source on the left and the player's render on the right, the same unscaled 700×880 crop at the default settings. They are the strongest frames found, not typical ones; the frame where it does not help is with the limits.
Flip swaps the picture between the two; 1:1 shows the captured files at one image pixel per screen pixel.
Rendering is not an export queue. Press D at any point and playback moves onto rendered frames a few seconds later, while the rest of the video renders behind you, nearest the playhead first. Seek anywhere: rendered frames play wherever they are, and elsewhere the original plays at once while the render moves there.
Split, wipe, side by side or 2 × 2, all on one timestamp, from a compare bar under the picture. Drag the divider across a face, hold the mouse to see the original in place, or step frames on either view.
Difference shows the change itself, amplified 1× to 32×. The loupe magnifies both at the same spot, and a setting changed while paused re-renders that frame, so you decide on the picture.
Save comparison image writes exactly what the picture shows as a PNG, with a footer recording the player version, video, frame, view, neural settings and runtime.
Renders stay on disk until you clear them or the drive falls below 20 GB free. Reopen a film weeks later and its whole-film render is still attached, as long as the source, the runtime and the neural settings still match.
Open a PNG, JPEG, BMP, TIFF or WebP and press D. The same model renders it, the same compare views show what changed, and it saves back as a PNG or JPEG at full size. Here the game’s smooth orange skin comes back with texture, a truer tone and light that falls across the face. Animated GIFs render frame by frame and keep their timing.
A video, a photo or an animated GIF: whatever you watch rendered, you can save. Save converted video writes the render you already have without rendering again; Export with DLSS stages (Ctrl+S) renders a new file. Both are under DLSS > Convert & export.
VideoMKV or MP4, of the whole video or a marked clip. MKV keeps the source audio, subtitles and chapters without re-encoding them.
GIFAn animated GIF renders frame by frame with its own timing, and saves as GIF, MP4 or MKV.
PhotoPNG, JPEG, BMP, TIFF or static WebP in; PNG or JPEG out, at the source size. The 2560×1440 screenshot above took 8 s from the command line on an RTX 4080 SUPER.
StagesSuper Resolution, neural rendering and frame generation at 2× to 5× the frame rate, any combination, in one file and in NVIDIA’s order.
ScriptedDLSSVideoPlayer.exe --render clip.mp4 writes clip-dlss.mkv without opening the player; the usage guide has the options.
Left outSubtitles travel as tracks and are never burned in. Image adjustments and playback upscaling are not baked in; the stages export is how a larger size goes into the file.
Details are in the usage guide. Nothing is uploaded: the work happens on your GPU and the cache stays on your disk.
--rendergh attestation verifyverify_package.ps1 checks every file against a pinned manifest before the first run
Stated as plainly as the features, and measured where a number is given.
HardwareAn NVIDIA RTX card and driver 610.47 or newer; older drivers are refused up front. Checked on an RTX 4080 SUPER and an RTX 5090. RTX 20 and 30 cards lack native FP8, should be several times slower, and are untested.
SpeedA live 1080p30 session costs 8.4 ms a frame on an RTX 5090 and 15.3 ms on an RTX 4080 SUPER. A whole 4K render on the 4080 SUPER ran at 14.8 fps, slower than a 30 fps film plays.
UpscalingDLSS Super Resolution scored below a plain bicubic upscale on all six video clips measured. It is built for games, so it is off by default and there for its look, not for detail.
RuntimeThe neural runtime is a modified, unsigned community build, not a signed NVIDIA package. Why it is unsigned.
ArtifactsDepth is estimated from the picture, and so is motion on cards without the optical flow engine. Expect some artifacts.
Low keyIn deep shadow the model can crush a face towards black and add speckle. The frame below is the plainest case found.
CacheA render started partway into a video is not joined to the rest, and below 20 GB free the least recently watched renders are deleted first.
v0.26.1, published 25 September 2026. Free, with no account and no installer: unzip it and run it. Every build is on the release page with its checksum.
The player and the pinned neural runtime. This is the one you want.
Download 327 MBdlss5-video-player-v0.26.1-win64.zip
The player alone, without the neural runtime. Build or supply your own.
Download 36 MBDLSSVideoPlayer-v0.26.1-core-win64.zip
MIT licensed. Read it, build it yourself, or check that the core zip was built from it.
View source 153 starsgithub.com/2600th/dlss5-video-player
neural-runtime/ next to the executable.DLSSVideoPlayer.exe.
It is open source, but its neural runtime is a modified community
build that nobody can sign, so decide with that in view. Each zip has a
SHA-256 beside it. The core zip is built by CI from this repository and
carries a signed build attestation (gh attestation verify);
the complete zip is assembled by the maintainer and has a checksum
only. Inside either, verify_package.ps1 checks every file
against a pinned manifest before the first run. Scan the runtime as you
would any unsigned software, or take the core zip and supply your own.
Because it has been modified. It includes NVIDIA's neural DLL patched to run on RTX 40 cards, which invalidates NVIDIA's signature, alongside community ReShade and RenoDX files that were never signed. Nobody is in a position to re-sign it, so it ships unsigned and says so. Its files are pinned by size and SHA-256, and the package verifier rejects any that have changed.
An NVIDIA RTX card and driver 610.47 or newer; with an older driver the player says so up front. It has been checked on an RTX 4080 SUPER and an RTX 5090. RTX 20 and 30 series cards lack native FP8 and should be several times slower; nobody has tested one yet. On any other GPU the start screen says "not an RTX GPU, so no DLSS", and playback there has not been tested.
You do not wait for it to start: playback joins the render a few seconds after you press D, and the rest fills in while you watch. A live 1080p30 session costs 8.4 ms a frame on an RTX 5090 and 15.3 ms on an RTX 4080 SUPER, both faster than the video plays; 1440p30 on the 5090 costs 15.4 ms. A whole 4K film on the 4080 SUPER rendered at 14.8 frames a second. After your first session the player warns you before a video it expects to fall behind on. The other limits are above.
No. Rendering runs on your GPU and the cache stays on your disk. The
player goes online to fetch a YouTube link you paste, and once a day
to ask GitHub whether there is a new release, which
[Updates] Enabled=0 in its settings file turns off.
The start screen also fetches the seven built-in trailers' thumbnails
from YouTube's image server, then again once a month, which
[Start] ThumbnailFetch=0 turns off.
It can, with DLSS Super Resolution, but that is off by default because on video it does not beat a plain scaler: on six clips upscaled from 960×540 to 1080p, bicubic scored higher every time. It is built for rendered games, not decoded footage, so turn it on for its look, not for detail. Neural rendering itself keeps the source resolution.
Yes. Open a PNG, JPEG, BMP, TIFF or static WebP and press
D: it renders once and stays on that frame, with the same
compare views as a video, and saves as a PNG or JPEG at its own size.
From the command line, --render turned a 2560×1440 game
screenshot into a PNG in about 8 seconds on an RTX 4080 SUPER.
Animated GIFs render frame by frame and keep their timing. The model
is built for rendered games, and game screenshots are what it has been
tried on; camera RAW and animated WebP are not supported.
Yes: PNG or JPEG for photos, GIF for animations, MP4 or MKV for video,
from the player or, from v0.26.0, from a script with
--render. MKV keeps the source audio, subtitles and
chapters without re-encoding them.
Export with DLSS stages writes Super Resolution, neural rendering and
frame generation into one file, in NVIDIA's order.