DLSS 5 Neural Rendering for Images & Video with Frame Generation, RTX Video Super Resolution, HDR, and Color Grading
TypeScript ai-upscalingdlssdlss-5dlss5dlssgframe-generationframe-interpolationimage-enhancementimage-upscalingneural-renderingnvidianvidia-dlss-5nvidia-rtxrtx-videortx-video-super-resolutionvideo-enhancementvideo-upscaling
Add files via upload
Merserkdb11368435c5
diff --git a/README.md b/README.mdindex 5a8fedc..b55e667 100644--- a/README.md+++ b/README.md@@ -2,7 +2,7 @@[](https://github.com/Merserk/dlss5-visual-enhancer/releases)             [](https://www.patreon.com/Merserk)-**Visual Enhancer** is a portable Windows application for AI-assisted image and video enhancement. Its configurable pipelines combine **NVIDIA DLSS 5 Neural Rendering** through the **Neuroframe Engine**, scaling, **DLSS Super Resolution**, **DLSS Ray Reconstruction**, **RTX Video Super Resolution**, image **Upscale** with **VisionNext**, **SharpNext**, or **RazorNext**, **ReShade**, coloring, sharpening, and film grain. Video processing also offers **RTX Video HDR**, **DLSS Frame Generation**, **FSR Frame Generation**, **Video Super Resolution**, **Video Frame Generation**, and **Upscale (Lightweight)** as separate stages.+**Visual Enhancer** is a portable Windows application for AI-assisted image and video enhancement. Its configurable pipelines combine **NVIDIA DLSS 5 Neural Rendering** through the **Neuroframe Engine**, scaling, **DLSS Super Resolution**, **DLSS Ray Reconstruction**, **RTX Video Super Resolution**, AI **Upscale** for images and video, **ReShade**, coloring, sharpening, and film grain. Video processing also offers **RTX Video HDR**, **DLSS Frame Generation**, **FSR Frame Generation**, **Video Super Resolution**, **Video Frame Generation**, and **Upscale (Lightweight)** as separate stages.<img width="1920" height="1080" alt="Main Cover_v2" src="https://github.com/user-attachments/assets/9248c946-391e-408c-b146-f7d8a6e9c695" />@@ -14,7 +14,7 @@2. Extract the complete ZIP archive to a folder.3. Run **`Visual Enhancer.exe`**.-AI image models are downloaded separately. Use the **Models** tab to download them, or enable **Upscale** or an AI-depth ReShade effect to start downloading the required model automatically. Internet access is needed for downloads; installed models run locally. A fresh package contains no previous outputs or session logs.+AI models are downloaded separately. Use the **Models** tab to download them, or enable **Upscale** or an AI-depth ReShade effect to start downloading the required model automatically. Internet access is needed for downloads; installed models run locally. A fresh package contains no previous outputs or session logs.## Examples@@ -41,7 +41,7 @@ The **Neural Rendering** workspace has separate **Image** and **Video** pipeline| **DLSS Super Resolution** | DLAA or AI upscaling with selectable quality modes and presets. Video includes optical-flow quality controls. || **DLSS Ray Reconstruction** | Reconstruct images and videos at native **1× resolution**. Video uses NVIDIA Optical Flow (NVOFA), with High, Medium, or Low optical-flow quality. || **RTX Super Resolution** | NVIDIA RTX Video enhancement and upscaling, with quality levels, scale factors, or custom dimensions. |-| **Upscale** | Image restoration with **VisionNext**, **SharpNext**, or **RazorNext**, using FP16 models through Windows ML. Select **1×, 2×, 4×, 6×, or 8×** output scale; **2×** is the default. Tile size or GPU memory controls depend on the selected model. |+| **Upscale** | AI upscaling with **VisionNext**, **RazorNext**, **PeakMagnetar**, **OptiMagnetar**, and **TinyMagnetar**. Images support **1×, 2×, 4×, 6×, and 8×**. The three Magnetar models also support **2× and 4×** video upscaling and fixed target resolutions. || **RTX Video HDR** | Convert SDR video to HDR with contrast, saturation, middle gray, peak luminance, and precision controls. || **DLSS Frame Generation** | Increase video frame rate with a 2×–8× multiplier or target FPS, automatic engine selection, and High, Medium, or Low optical-flow quality. || **FSR Frame Generation** | Generate intermediate video frames using DirectX 12 and AMD FidelityFX Optical Flow, with a 2×–8× multiplier or target FPS. FSR processing runs on the GPU without CUDA or NVIDIA Optical Flow. |@@ -57,15 +57,18 @@ Ray Reconstruction adapts a model designed for ray-traced rendering to importedVideo-only effects appear as individual cards in the Video pipeline. Stages run from top to bottom; the application checks hardware support, dimensions, frame rates, and HDR compatibility before processing. Choose one frame-generation provider per pipeline: **DLSS**, **FSR**, or **Video Frame Generation**. DLSS and FSR offer **Multiplier** and **Target FPS** modes, with fixed targets from **23.976 to 480 FPS**; multiplier mode retains the source's fractional frame rate. The **Frame Interpolation** workspace also lets you choose between DLSS and FSR. ReShade includes a bundled preset, shader and texture search paths, precision and HDR controls, and optional AI depth generation; RTGI shaders are not bundled.-**Upscale** is an Image-mode effect; it is separate from the video-only **Upscale (Lightweight)** stage. VisionNext and SharpNext offer automatic or manual tile sizes. RazorNext instead offers an automatic or selectable GPU memory budget; lower budgets can reduce GPU memory use at the cost of processing time and additional system memory. RazorNext can require substantial GPU and system memory. Image model sessions can remain loaded briefly between operations and are released after inactivity or when memory is needed; cached image results can also be reused.+**Upscale** and **Upscale (Lightweight)** are separate stages. VisionNext offers automatic or manual tile sizes. The Magnetar models handle tiling automatically. Fixed video targets preserve the whole picture with black bars when aspect ratios differ. Export settings disable codecs that cannot handle the selected dimensions. RazorNext instead offers an automatic or selectable GPU memory budget; lower budgets can reduce GPU memory use at the cost of processing time and additional system memory. RazorNext can require substantial GPU and system memory. Image model sessions can remain loaded briefly between operations and are released after inactivity or when memory is needed; cached image results can also be reused.++**SharpNext has been discontinued** and is no longer available in the Upscale model selector.### Workflow and preview- **Comparison:** Split, 2-Up, and Output views, zoom, fit-to-view, a shared video timeline, and Focus Preview (`Ctrl+Shift+F`).+- **Compare workspace:** compare images and videos with Side-by-Side, Split View, Cycle, and Difference modes. Supported videos use GPU-resident playback at the original source resolution. Save sessions and export comparisons.- **Previews:** realtime updates for supported changes, full-resolution image previews, and short video previews from the selected position. Cached results reduce repeated processing.-- **Models:** manage **VisionNext**, **SharpNext**, and **RazorNext** under **Upscale**, and **MoGe-2** under **ReShade**. Download, cancel, resume interrupted transfers, or remove models with confirmation. Downloads verify file hashes and retry connection failures. The **Models** tab shows an animated progress underline during background downloads, including downloads started by enabling an effect.+- **Models:** manage **VisionNext**, **RazorNext**, **PeakMagnetar**, **OptiMagnetar**, and **TinyMagnetar** under **Upscale**, and **MoGe-2** under **ReShade**. Download, cancel, resume interrupted transfers, or remove models with confirmation. Downloads verify file hashes and retry connection failures. The **Models** tab shows an animated progress underline during background downloads, including downloads started by enabling an effect.- **Batches:** add files or folders, drag and drop, paste clipboard media, track per-file progress, pause or cancel exports, and retry failed items. **Show input in Explorer** and **Show output in Explorer** select the corresponding file in File Explorer.-- **Export:** choose the format, codec, container, quality, output folder, and naming mode. Existing files are not silently overwritten.+- **Export:** choose the format, codec, container, quality, output folder, and naming mode. Auto naming uses the `_visual_enhancer` suffix and adds `_2`, `_3`, and so on when needed.- **HDR:** preserve or export 10-bit HDR with a compatible pipeline and codec. Place SDR-only effects before HDR conversion.- **Presets:** save and restore processing settings, stage order, and enabled effects as JSON.@@ -74,9 +77,9 @@ Video-only effects appear as individual cards in the Video pipeline. Stages run| Type | Input | Output || --- | --- | --- || Images | PNG, JPEG, JPEG XL, WebP, AVIF, TIFF, BMP, GIF, SVG, HEIF/HEIC, and many camera RAW formats | PNG, JPEG, JPEG XL, WebP, AVIF, TIFF; 16-bit export for PNG, TIFF, and JPEG XL |-| Video | MP4, MKV, MOV, AVI, WebM, M4V, TS/MTS/M2TS, WMV, FLV | H.264, H.265, AV1, ProRes Proxy/HQ, FFV1 Lossless RGB 10-bit; compatible MP4, MKV, or MOV containers |+| Video | MP4, MKV, MOV, AVI, WebM, M4V, TS/MTS/M2TS, WMV, FLV | H.264, H.265, AV1, VP9, ProRes Proxy/HQ, FFV1 Lossless RGB 10-bit; compatible MP4, MKV, MOV, or WebM containers |-Image processing applies supported color profiles and EXIF orientation, retains supported metadata, and preserves transparency where the output format allows it. Animated or multipage images use the first frame/page. Video exports retain audio, metadata, and chapters where supported; subtitle preservation depends on the processing path and container.+Image processing applies supported color profiles and EXIF orientation, retains supported metadata, and preserves transparency where the output format allows it. Animated or multipage images use the first frame/page. Video exports retain audio, metadata, and chapters where supported. VP9 defaults to WebM, which also supports AV1. WebM preserves compatible audio, converts other audio to Opus, and converts supported text subtitles to WebVTT.The application accepts image dimensions up to **32,768 pixels per side**, subject to available memory and the limits of the selected effects and file format.@@ -84,7 +87,7 @@ The application accepts image dimensions up to **32,768 pixels per side**, subje- **Windows:** Windows 11, 64-bit.- **GPU and driver:** NVIDIA GeForce RTX with a compatible NVIDIA driver for DLSS, RTX Video, and NVIDIA Video Effects. Super Resolution and Lightweight Upscale support RTX 20, 30, 40, and 50 series; Neural Rendering availability is checked on the selected GPU. FSR Frame Generation has its own DirectX GPU selection and compatibility checks.-- **Image Upscale:** a DirectX 12 hardware GPU compatible with DirectML, with enough GPU and system memory for the selected model and output size.+- **AI Upscale:** a DirectX 12 hardware GPU compatible with DirectML, with enough GPU and system memory for the selected model and output size.- **Ray Reconstruction video:** NVIDIA Optical Flow support must be available through the installed NVIDIA driver. Images do not require optical flow.- **Frame generation:** DLSS Frame Generation and NVIDIA Video Frame Generation require supported RTX 40 or 50 series hardware. Enable **Hardware-accelerated GPU scheduling** in Windows for DLSS Frame Generation. FSR Frame Generation uses compatible DirectX 12 hardware and AMD FidelityFX Optical Flow; adapter support is checked before processing.- **Video filtering:** a working Vulkan-capable GPU and driver for Vulkan filter stages. NVIDIA, AMD, or Intel can handle this path. FSR Frame Generation uses DirectX 12 for its own processing.@@ -93,7 +96,7 @@ NVIDIA NVENC export requires an encoder supported by the selected RTX GPU. AV1 h## Settings-**Settings** contains interface language, **AI Processing GPU (Tensor Cores / NGX)** for NVIDIA processing, **AI Processing GPU (DirectX 12 / FidelityFX)** for FSR Frame Generation, a separate **Upscale → GPU** selector and **AI Inferencing Framework** for image Upscale, **Decoding / Encoding Device**, **Realtime Preview**, and **Processing Cache**. The processing GPU selections are independent of the general video device setting. Image Upscale can use automatic GPU provider selection or DirectML; first-time model preparation can take longer, and compiled models can be reused later. FSR's **Auto** selection prioritizes a compatible discrete GPU using Windows' high-performance adapter preference. **Clean Cache** releases reclaimable preview, shader, and processing cache data. Preset import and export are available from the header menu.+**Settings** contains interface language, **AI Processing GPU (Tensor Cores / NGX)** for NVIDIA processing, **AI Processing GPU (DirectX 12 / FidelityFX)** for FSR Frame Generation, a separate **Upscale → GPU** selector and **AI Inferencing Framework** for AI Upscale, **Decoding / Encoding Device**, **Realtime Preview**, and **Processing Cache**. The processing GPU selections are independent of the general video device setting. AI Upscale can use automatic GPU provider selection or DirectML; first-time model preparation can take longer, and compiled models can be reused later. FSR's **Auto** selection prioritizes a compatible discrete GPU using Windows' high-performance adapter preference. **Clean Cache** releases reclaimable preview, shader, and processing cache data. Preset import and export are available from the header menu.Use the processing cards for individual effects and **Export Settings** for image format, bit depth, video codec, container, quality, and **Preserve / Export 10-bit HDR**. The Export dialog selects the destination and filename. Saved settings, caches, and downloaded models use Windows per-user storage; replacing the application folder does not reset them. The default output folder is `outputs/` beside the executables, and session logs are stored in `logs/`; these folders are created when the application starts. Temporary processing files use application-managed temporary storage.@@ -123,7 +126,7 @@ The interface supports the following languages, with language changes applied im| `render` | Neural Rendering, or the complete enabled pipeline from a JSON preset. || `reconstruct` | DLSS Ray Reconstruction for images and videos at native 1× resolution. || `reshade` | Process images or videos with a ReShade `.ini` preset. |-| `upscale` | RTX or DLSS Super Resolution; VisionNext, SharpNext, or RazorNext for images. |+| `upscale` | RTX or DLSS Super Resolution, VisionNext and RazorNext for images, or Magnetar models for images and video. || `interpolate` | DLSS or FSR Frame Generation with a multiplier or target FPS. || `devices` | List NVIDIA processing GPUs and DirectX adapters for FSR. |@@ -132,7 +135,7 @@ The interface supports the following languages, with language changes applied im.\VE_CLI.exe reconstruct "photo.png" -o "results\photo-rr.png".\VE_CLI.exe reshade "photo.png" --reshade-preset "look.ini" -o "results\photo-reshade.png".\VE_CLI.exe upscale "photo.png" --engine visionnext --scale 2 --tile auto -o "results\photo-visionnext.png"-.\VE_CLI.exe upscale "photo.png" --engine sharpnext --scale 2 --tile auto -o "results\photo-sharpnext.png"+.\VE_CLI.exe upscale "photo.png" --engine peakmagnetar --scale 2 -o "results\photo-peakmagnetar.png".\VE_CLI.exe upscale "clip.mp4" --scale 2 -o "results".\VE_CLI.exe interpolate "clip.mp4" --fps 60 -o "results".\VE_CLI.exe interpolate "clip.mp4" --interpolator fsr3 --fsr-gpu auto --fps 60 --codec hevc -o "results\clip-fsr.mp4"@@ -142,7 +145,7 @@ The interface supports the following languages, with language changes applied imUse `--help` for options, `--recursive` for subfolders, `--dry-run` to inspect a job, and `--json --quiet` for scripting. Use `reconstruct --help` for RR options; RR always processes at 1× and does not accept DLSS mode or RR preset options. `interpolate` defaults to DLSS; select FSR with `--interpolator fsr3`, then choose either `--fps` or an integer `--multiplier` from 2 through 8. `--fsr-gpu` accepts `auto`, a DirectX adapter index, or a saved DXGI ID from `devices --json`, independently of the NVIDIA `--gpu` option. ReShade accepts `--reshade-preset`, shader and texture search paths, and depth controls; use `reshade --help` for details. Presets exported from the application include advanced effect and export settings. Without `-o`, output goes to `outputs/` beside the executables.-For AI image Upscale, select `--engine visionnext`, `--engine sharpnext`, or `--engine razornext`. Each accepts `--scale 1|2|4|6|8`; `--tile` accepts `auto|256|512` for VisionNext or `auto|128|256` for SharpNext. RazorNext uses `--memory-budget auto|unlimited|12|8|6` instead of manual tiling. Use `--upscale-gpu auto` or a saved DirectX adapter ID to select the Upscale GPU independently of NVIDIA processing. Required models download automatically if they are missing. ReShade's AI-depth model is selected with `--reshade-depth moge-2-vitb`; `auto` downloads it only when the preset requires depth.+For AI image Upscale, select `--engine visionnext`, `--engine razornext`, `--engine peakmagnetar`, `--engine optimagnetar`, or `--engine tinymagnetar`. Each accepts `--scale 1|2|4|6|8`. VisionNext supports `--tile auto|256|512`. RazorNext uses `--memory-budget auto|unlimited|12|8|6` instead of manual tiling. Use `--upscale-gpu auto` or a saved DirectX adapter ID to select the Upscale GPU independently of NVIDIA processing. Required models download automatically if they are missing. ReShade's AI-depth model is selected with `--reshade-depth moge-2-vitb`; `auto` downloads it only when the preset requires depth.## Source Code Availability@@ -154,6 +157,6 @@ Source code may not be updated with every release. Updates will be published perProject-owned Visual Enhancer material is distributed under the **Merserk Source License**. Its terms permit personal, professional, and commercial work and commercial use of outputs, subject to applicable third-party terms. Redistribution, mirroring, repackaging, rebranding, resale, sublicensing, publishing modified builds, and similar redistribution require prior written permission except where the license states otherwise. See [LICENSE](LICENSE) for the complete terms.-Third-party components and downloaded models remain subject to their own licenses and are not relicensed under the Merserk Source License. **SharpNext incorporates ResShift under the S-Lab License, which permits non-commercial use; LPNSR's MIT license does not remove that restriction.** See [THIRD-PARTY-NOTICES](THIRD-PARTY-NOTICES) and the model package's `LICENSE.txt` for the applicable terms.+Third-party components and downloaded models remain subject to their own licenses and are not relicensed under the Merserk Source License. See [THIRD-PARTY-NOTICES](THIRD-PARTY-NOTICES) and the model package's `LICENSE.txt` for the applicable terms.Visual Enhancer is an independent community project and is not affiliated with, sponsored by, or endorsed by NVIDIA or AMD. NVIDIA, GeForce RTX, DLSS, and RTX Video are trademarks and/or registered trademarks of NVIDIA Corporation. AMD and FidelityFX are trademarks of Advanced Micro Devices, Inc.