DLSS 5 Visual Enhancer: a Windows Gradio front end for Neural Rendering, RTX VSR and DLSSG frame interpolation

What DLSS 5 Visual Enhancer actually wraps

Three NVIDIA technologies that normally live in different places get one interface here. The first is DLSS 5 Neural Rendering, run through what the README calls the Neuroframe Engine, a self-contained engine built around the NVIDIA NGX/DLSS runtime. The README states that Neural Rendering no longer depends on an external graphics injector or add-on. The second is NVIDIA RTX Video Super Resolution and RTX Video HDR, exposed as a separate Upscale workflow for images and videos. The third is NVIDIA DLSS Frame Generation, used for video frame interpolation with selectable output rates from 23.976 to 480 FPS.

The audience is narrow and specific: people on Windows with an NVIDIA RTX card who already have image or video files they want processed, and who would rather click through a local Gradio page than assemble a pipeline. The README describes a portable application, and the repository layout backs that up: app.py, bin/, src/ and start.bat at the top level. There is no server component, no account, and no cloud round trip described anywhere. The project is explicit that it is an independent community project and is not affiliated with, sponsored by, or endorsed by NVIDIA. Treat that line as load-bearing when you evaluate support expectations.

How the Neuroframe Engine moves a frame through the pipeline

For video, the README describes a decode, DLSS, encode path. When the VRAM processing path is selected and the output codec is an NVIDIA NVENC variant, supported Neural Rendering video jobs can keep more of that path on the GPU, which the README frames as reducing unnecessary transfers through system memory. The alternative is a RAM staging path, selectable from Settings. That is a real architectural fork, not a cosmetic toggle: one path trades VRAM for fewer copies, the other trades copies for headroom.

Neural Rendering itself is not a single pass. The Multi Pass control runs it from 1 to 4 passes, and the README describes additional passes as a way to get stronger enhancement or refinement. On top of that sits a composition layer that decides how much of the Neural Rendering result is blended back with the original: NR Color Strength, Tone Preservation, Face/Skin Protection, Grain Preservation and Mask Feather. The Detail-Only preset is just a named combination of two of those values, with NR Color Strength at 0 and Tone Preservation at 1.

Where Neural Rendering applies is controlled separately. Automatic Mask handles it automatically, or you can upload a custom image mask. One detail worth noting: the same selected mask is shared between Image, Video and Live for the current application session, so a mask you set for a still will follow you into a video job until you change it. Processing resolution is also decoupled from output size. The Neural Rendering scale offers Source, 75%, 50% and 25%, and the README is clear that these control the resolution processed by Neural Rendering while enlargement is handled separately by Upscale mode. Those are two different stages, and conflating them is the easiest way to misread the settings panel.

Installing it and running a first image through Neural Rendering

The README gives three installation steps and no build step. Download the latest release ZIP from the releases page, unpack it, and run start.bat. There is no installer, no package manager entry and no dependency list in the README, which fits the portable framing: the bin/ directory is presumably where the runtime lives, though the README does not document its contents.

start.bat

Running start.bat should bring up the local Gradio interface. The README does not state a port number, so read the console output after launch rather than guessing a URL. Once the page is open, the first real use is a single image. Pick the Image workflow, point it at one file, and leave the defaults alone before you touch anything else. The README lists common Pillow formats plus HEIF/HEIC, SVG and many camera RAW formats as inputs, and PNG, JPEG, WebP, AVIF or TIFF as outputs. It also notes that EXIF orientation is applied, embedded ICC input is converted to sRGB, supported EXIF/DPI/XMP metadata is retained, and alpha is preserved except when JPEG composites transparency over white. Animated and multipage sources use the first frame or page, which matters if you feed it a GIF and expect motion back. Use the Preview action before committing, then raise Multi Pass to 2 or 3 and compare. If you want the source color untouched, switch on Detail-Only rather than hand-tuning NR Color Strength.

Where it breaks: HDR, codecs and the missing documentation

The HDR story has a hole in it. HDR Mode can preserve HDR with 10-bit output, but only when H.265, AV1 or ProRes Proxy is selected. H.264 remains 8-bit SDR. If your delivery target is H.264, HDR Mode buys you nothing, and the README does not offer a workaround.

The GPU pipeline has a similar condition attached. Keeping decode, DLSS and encode on the GPU requires both the VRAM processing path and an NVIDIA NVENC output codec. ProRes Proxy is CPU-based, so choosing it pulls the job off that path by design. The README does not quantify the cost of either path, so the only way to know whether the VRAM path is faster on your card is to run the same clip both ways.

The larger problem is documentation depth. The README is a feature list. It does not document rollback, it does not document error messages, it does not state a port for the Gradio interface, and it does not list hardware requirements beyond the Windows and RTX badges. There is no CLI described anywhere, which means batch work happens by pointing the Input field at a folder rather than by scripting. If your workflow needs a headless job in CI, this is the wrong tool, and no amount of configuration will change that. Live mode adds its own constraint: the README states that scale and source/playback configuration take effect on the next Start, so those two settings cannot be changed mid-session the way NR Intensity or the composition controls can.

Neural Rendering versus RTX Video Super Resolution: two different jobs

The most useful comparison is internal to the project. Neural Rendering and Upscale are separate workflows with separate controls, and picking the wrong one wastes time.

Upscale for images uses NVIDIA RTX Video Super Resolution with VSR quality levels 1 to 4 and 1x to 4x scaling or custom dimensions, plus aspect-ratio locking and metadata preservation. Upscale for videos goes further: it can apply RTX Video Super Resolution, RTX Video HDR, or both in the same workflow, and RTX Video HDR exposes adjustable contrast, saturation, middle gray, peak luminance up to 2000 nits, and selectable HDR processing precision. That is a scaling and tone-mapping tool. Its job is to make a frame bigger or brighter.

Neural Rendering is a detail and composition tool. It changes structure and tone through NR Style, NR Intensity, Local Tone Strength, Local Structure Strength and Skin Structure Strength, with masks deciding where those changes land. Shimmer Suppression exists specifically for video and Live, described as temporal stabilization for reducing unstable fine detail and shimmer between frames. If your source is already the right size and you want it to look different, that is Neural Rendering. If it is the wrong size, or SDR when you need HDR, that is Upscale. The README's own separation of processing scale from enlargement says the same thing in configuration terms.

Live playback and the settings that survive a restart

Live mode applies Neural Rendering during playback of local videos, direct network streams, YouTube and Twitch, and the README states that effect changes can be applied while the session is running without restarting playback processing. The list of live-adjustable controls is long: NR Style, NR Intensity, NR Passes, the tone and structure controls, the composition controls, Shimmer Suppression, Automatic Mask, Custom NR Mask and the Processing Engine Path.

The exceptions are what you should plan around. Scale and source/playback configuration take effect on the next Start. Playback configuration itself covers source quality, processing height from 480p to 2160p, Auto/Source/60/30/24 FPS modes, 1/2/4-second segments, adjustable playback buffering, and optional automatic playback in the bundled MPV player. If you are tuning a stream, settle the height and segment length first, then tune the Neural Rendering controls live. Doing it the other way round means restarting the session to test a resolution change.

Live also inherits the shared mask behaviour. A custom NR Mask selected in the Image or Video workflow is the same mask Live uses for the current session, so a mask drawn for a portrait still will silently apply to whatever you play next.

Licence, releases and what maintenance costs you

The repository is MIT licensed, with the LICENSE file at the top level. MIT is permissive: it lets you use, modify and redistribute the code, including commercially, provided the copyright notice and permission notice are kept. It also means the author offers no warranty, which is consistent with a project whose README carries an explicit non-affiliation disclaimer with NVIDIA. None of that touches the NVIDIA components the app drives, and the README says nothing about the terms attached to the NGX/DLSS runtime or the NVENC encoders. If you plan to ship output commercially, that is a question for your own counsel, not something the README answers.

Release cadence is visible and fast. v7.0 arrived on 2026-09-05, v8.0 on 2026-09-11, and v9.0 on 2026-09-13, with the last push to the repository on 2026-09-14. Four days between v8.0 and v9.0 is a short window, and the README describes v8.0 composition controls as expanded, which implies the settings surface is still moving. Budget for re-reading the release notes before each upgrade, and expect that a settings panel you learned in v8.0 may gain controls in v9.0. The upgrade mechanism itself is the same as installation: download the new ZIP and unpack it. The README does not document an in-place upgrade path, so treat each release as a fresh unpack.

Editorial conclusion

Adopt it if you are on Windows with an RTX GPU and you want Neural Rendering, RTX Video Super Resolution and DLSS Frame Generation behind one Gradio UI instead of three separate tools, and if you accept that the README is the only documentation. Do not adopt it if you need Linux, a scriptable CLI, or a support contract: the repository ships app.py, bin/, src/ and start.bat, and nothing else. Before you commit a batch job, verify two things on a single clip: that your GPU has the VRAM headroom for the VRAM processing path, and that your chosen output codec is one the README lists as GPU-capable, because ProRes Proxy is CPU-based and H.264 stays 8-bit SDR even in HDR Mode.

Frequently asked questions

Official sources

1. Issues
2. License: MIT
3. Merserk/dlss5-visual-enhancer on GitHub
4. README
5. Releases