About this project

DLSS 5 Video Player is a free, open-source Windows player that sends video, photos or GIFs through NVIDIA's DLSS 5 neural renderer and lets you inspect the result against the source frame by frame. It is a community project, not an NVIDIA product, and the neural runtime it ships is a modified, unsigned community build. What it does - Renders while you watch: playback moves onto rendered frames a few seconds after you press D, while the rest of the video fills in behind you, nearest first. - Seeks anywhere: rendered frames play wherever they exist; elsewhere the original plays while the render catches up. - Comparison views: Split, Wipe, Difference (amplified change), Side by side and 2 x 2, plus a Mix slider, zoom to 8x, a 4x loupe, a mask limiting DLSS 5 to part of the frame, and press-and-hold for the original. Switching views does not move the playhead. - RTX Video Super Resolution of the original can be shown as its own view and in the 2 x 2 (R, Shift+R), in the complete download or a build made with the RTX Video SDK. - Save comparison image writes exactly what the picture shows, with a footer recording video, frame, view, neural settings and runtime. - Subtitles (SRT, ASS, WebVTT, PGS, VobSub) are drawn over the render so the model never warps text. HDR10 and HLG are tone mapped to SDR for the model; on an HDR display the original still shows in HDR. - Renders are cached and reused while source, runtime and settings match, at Standard, High (10-bit) or Lossless quality. A render report shows flicker, grain and colour shift added by a render. - Exports: PNG, JPEG, GIF, MP4 or MKV; MKV keeps source audio, subtitles and chapters without re-encoding. Export with DLSS stages (Ctrl+S) combines Super Resolution, neural rendering and frame generation in NVIDIA's order, and a --render command-line mode does the same. Generate frames writes a copy at 2x to 5x the frame rate. - Neural settings can be changed while paused, re-rendering that frame; a processing scale runs the model at 75% or 50% of the source for speed. - Player features: AC-3, E-AC-3 and DTS passthrough, Windows media controls with taskbar thumbnail buttons, and seven official game trailers bundled as test material. Requirements and limits - Needs an NVIDIA RTX card and driver 610.47 or newer; older drivers refuse neural rendering and the player says so up front. - Tested on an RTX 4080 SUPER (0.26.0) and RTX 5090 (0.20.0). RTX 20 and 30 series cards lack native FP8 and are expected to be several times slower; the README states nobody has tested one yet. - Stated live-session cost: about 8.4 ms per frame at 1080p30 and 15.4 ms at 1440p30 on an RTX 5090, and 15.3 ms at 1080p30 on an RTX 4080 SUPER. 4K depends on the file. - Depth and, on cards without the optical flow engine, motion are estimated from the picture, so artifacts are expected. - The README reports DLSS Super Resolution scoring lower than bicubic on six test clips, and advises using it for its look rather than detail. - Frame generation writes a new file and needs time and disk space; a stream must be copied locally first. - Save converted video copies the cached render as-is (8-bit at Standard, 10-bit at High or Lossless) without adjustments, upscaling or HDR. - Not supported: burning subtitles into an export, a render queue, resuming an interrupted render after restart. - YouTube support is limited to public, non-DRM videos with no login; age-restricted videos may arrive as 640x360. - The cache keeps renders until the drive falls below 20 GB free, then deletes least recently used first. - Partial renders are not joined, so a session started partway through a video leaves separate renders and reopening the video renders it again. - Network use: a daily GitHub release check and start-screen thumbnail fetches from i.ytimg.com; both can be disabled in DLSSVideoPlayer.ini. Download and first run Version 0.26.0 (2026-09-25) is offered as a 327 MB complete zip with the player, neural runtime and RTX VSR, or a 35 MB core zip with the player only. Each zip has a .sha256; only the core zip has a build attestation, because CI cannot fetch the neural runtime. A verify_package.ps1 script checks unpacked files against the package manifest and shows which binaries are signed. Unzip into a new empty folder, keep neural-runtime/ next to the exe, run DLSSVideoPlayer.exe, open a file or paste a public YouTube link, then press D. Controls are keyboard-driven: Ctrl+O open file, Ctrl+L YouTube URL, Space play/pause, D neural rendering, C/Shift+C compare modes, X swap sides, [ and ] change Mix, R/Shift+R RTX VSR, Z/Shift+Z zoom, L loupe, Ctrl+Shift+S save view as PNG, V subtitles, ? or F1 all shortcuts, F/Shift+F/Ctrl+R render one frame/four seconds/marked clip, Ctrl+N neural settings, Ctrl+S export with DLSS stages, Ctrl+E image adjustments, F11 fullscreen. Settings persist between launches. Building and license The project is C++20 with Win32, Direct3D 12, FFmpeg and NVIDIA NGX. Documentation covers building and testing, runtime setup, architecture and a technical overview, plus hardware verification records. The project source is MIT; third-party binaries, game footage and trademarks keep their own terms. It began from DLSS 5 Video Player by Jessica Natalia Mods.