Merserk/dlss5-visual-enhancer v23.0

Visual Enhancer v23.0

on GitHub

Changelog

• Added Style Transfer workspace, introducing a dedicated tab to match the color grading, tone curves, highlights, micro-contrast, texture, and finish of any reference image onto content images (supporting RAW, HEIF, AVIF, JPEG XL, and standard formats up to 64 megapixels with automatic letterboxing detection), complete with Original/Output switching, Fit/100% zoom controls, Intensity/Color Grade/Post-processing (0–200%) sliders, full-resolution PNG rendering, and Adobe After Effects Preset (.ffx) export with embedded 3D color LUTs.
• Added AI Denoising (ClearMagnetar), bringing a dedicated, reorderable native-resolution denoising stage to the top of both Image and Video Neural Rendering pipelines - complete with Noise Adaptation for suppressing real-world grain and compression artifacts, high bit-depth SDR support, and full alpha transparency and ICC color profile preservation.
• Added AI Deblurring (FocusPulsar), introducing a dedicated, reorderable native-resolution restoration stage to recover sharp detail from motion and focus blur in both Image and Video workflows while preserving transparency, high bit-depth SDR color, and ICC profiles.
• Upgraded AI processing engine to NeuroFrame Runtime, replacing the older Windows ML backend with a dedicated FP16 PyTorch GPU runtime for NVIDIA CUDA 13.2, AMD ROCm / HIP, and Intel XPU - delivering optimized GPU acceleration, warm model caching between renders, multi-GPU selection, updated AI Inferencing Framework options (Auto, CUDA, XPU, ROCm / HIP), and automatic migration of previous settings.
• Replaced OptiMagnetar and TinyMagnetar with OptiPulsar, unifying fast AI upscaling across Image, Video, and CLI workflows while automatically migrating existing settings and saved presets to OptiPulsar.
• Expanded Video AI Upscaling, unlocking RazorNext and VisionNext for Video processing alongside PeakMagnetar and OptiPulsar in both the desktop app and CLI - delivering streaming temporal-window video reconstruction, automatic scene-cut resets, temporal still preview integration, and seamless compatibility with NVIDIA DLSS and AMD FSR Frame Generation.
• Improved GPU memory and tiling automation, replacing manual RazorNext GPU Memory and VisionNext Tile Size controls with smart automatic VRAM budgeting, dynamic tile-size fallback, and host/block offloading so demanding 4K and 8K exports complete reliably without manual tuning.
• Redesigned Models manager and runtime orchestration, organizing downloads into Upscale, Denoising, Deblurring, ReShade, and Runtime categories with package sizes displayed for every entry, automatic NeuroFrame Runtime installation ahead of model downloads, automatic resumption of queued previews and renders once installations finish, streamlined model removal UI, verified offline library migration, and customizable model-download-token.txt support.
• Upgraded ReShade (beta) AI Depth independence and responsiveness, moving MoGe-2 depth estimation onto a dedicated ONNX Runtime 1.24.4 + DirectML 1.15.4 pipeline sharing the active D3D12 queue with compiled depth-session caching to eliminate repeated recompilation delays when adjusting effects and improve GPU cleanup on shutdown.
• Improved pipeline throughput, preview caching, and queue telemetry, overlapping CPU software video encoding alongside PyTorch inference, isolating preview cache invalidation per mode and stage so export tweaks no longer discard active previews, reporting precise sub-1.0 FPS speeds (down to <0.01 FPS), burst-accurate throughput excluding initial setup time, live countdown ETAs, clearer localized Cancelled and Skipped queue statuses, safe active job cancellation on item removal, and Style Transfer coordination during cache cleanups.
• Fixed window chrome, zoom previews, UI layout, and 11-language localizations, restoring reliable Windows taskbar and Alt-Tab application icons on frameless windows, fixing native 100% zoom image preview requests, refining narrow-window header tab spacing and Compare empty-state messaging, and updating localized strings across all 11 supported languages.
• Updated bundled codecs, metadata tools, and core libraries, refreshing FFmpeg, JPEG XL, HEIF, ExifTool, and underlying image and compression libraries for improved format compatibility, reliability, and performance.