Skip to content

Upscaling & Resolution

Visual Enhancer separates upscaling into four specialized stages so you can match the right engine to photography, CGI, animation, low-bitrate streaming video, or fast resizing.

For lightweight NVIDIA VFX video upscaling cards (Video Super Resolution and Upscale (Lightweight)), see Ray Reconstruction & Motion.

Bilinear 2×: Before and after comparison between bilinear 2x scaling and VisionNext 2x AI upscaling on a brass astrolabe, maidenhair fern, and woven burlapVisionNext 2×: Before and after comparison between bilinear 2x scaling and VisionNext 2x AI upscaling on a brass astrolabe, maidenhair fern, and woven burlap
2× AI upscale via VisionNext reconstructing crisp degree markings, wet fern serrations, and burlap weave compared with a standard bilinear upscale.
Stage Best For Hardware / Runtime Scale Range Video SDR Rule
DLSS Super Resolution Crisp, fast GPU upscaling (DLAA–3×) with clean edges & motion NVIDIA RTX (>= 7.5) DLAA (1×) – 3× SDR required before RTX Video HDR
Upscale (Image: VisionNext, RazorNext, PeakMagnetar, OptiPulsar) High-detail photography, textures, & alpha/ICC artwork (VisionNext/RazorNext) NeuroFrame Runtime (CUDA, XPU, ROCm / HIP) 1×, 2×, 4×, 6×, 8× Image workflow
Upscale (Video: PeakMagnetar, OptiPulsar, RazorNext, VisionNext) High-fidelity temporal/spatial video synthesis & exact target sizes NeuroFrame Runtime (CUDA, XPU, ROCm / HIP) 2×, 4× or Target (480p–8K) SDR required before RTX Video HDR
RTX Video Super Resolution Compressed web video, streaming rips, & de-blocking upscales NVIDIA RTX (>= 7.5) 1×, 1.5×, 2×, 3×, 4× or Custom px SDR required before RTX Video HDR
Scaling Deterministic resize, supersampled downscaling, & nearest-neighbor pixel art Vulkan / CPU 25% – 1600% SDR required before RTX Video HDR (Video)

DLSS Super Resolution uses NVIDIA NGX temporal/spatial reconstruction (1× DLAA up to 3×). It is mutually exclusive with DLSS Ray Reconstruction.

  • Dimensions: Minimum input 32 × 32 px; maximum output 16,384 × 16,384 px. Video outputs automatically round to even pixel dimensions.
Control Options Default Notes
Mode DLAA (1×), Quality (1.5×), Balanced (1.724×), Performance (2×), Ultra Performance (3×) Quality DLAA refines anti-aliasing at native resolution; Performance/Ultra Performance maximize scale
Preset Default, J, K, L, M Default Selects NGX transformer/render preset behavior; switch presets and inspect Preview for fine texture differences
Optical Flow Quality High, Medium, Low (Video only) High Controls temporal motion-vector grid resolution on video

Configure hardware backend under Settings > Upscale (GPU + AI Inferencing Framework: Auto, CUDA, XPU, or ROCm / HIP) and download models from the Models view.

Model Best Use Case Alpha & ICC Support Underlying Synthesis Passes
VisionNext (default) General photography, fine detail & illustrations Yes (RGBA + ICC preserved) 4× AI synthesis + Lanczos resample
RazorNext Ultra-crisp edges, graphics, & intricate textures Yes (RGBA + ICC preserved) 4× AI synthesis + Lanczos resample
PeakMagnetar Clean photoreal restoration & natural gradients Yes (RGBA + ICC preserved; SDR only) 2× (for 1×/2×) or 4× (for 4×/6×/8×) + resample
OptiPulsar Fast, balanced photoreal detail enhancement Yes (RGBA + ICC preserved; SDR only) 2× (for 1×/2×) or 4× (for 4×/6×/8×) + resample
  • Upscale Factor: 1×, 2× (default), 4×, 6×, or 8×.
Model Execution Behavior Best For
PeakMagnetar (default) Frame-by-frame high-detail AI synthesis Clean camera footage & low-noise cinema clips
OptiPulsar Fast, clean per-frame enhancement Everyday upscaling when throughput matters
RazorNext Temporal multi-frame inference with automatic lookahead window Reconstructing moving detail cleanly with lower temporal shimmer
VisionNext High-detail 4× neural synthesis Detailed animations, graphics, and heavily upscaled sources
Control Options Default Notes
Upscale mode Multiplier, Target Resolution Multiplier Choose scale factor or exact standard frame size
Scale factor 2×, 4× 2× Active when Upscale mode = Multiplier
Target Resolution 640×480, 1280×720, 1920×1080, 2560×1440, 3840×2160, 7680×4320 1920×1080 Preserves square pixels with letterboxing if aspect ratio differs; chains 2×/4× passes + final fit

Both videos play silently in a synchronized loop. Drag or use the arrow keys to compare. Press Space to play or pause, or Escape to pause.

Video Upscale using RazorNext in Multiplier mode (2×, 854×480 → 1708×960) reconstructing overhanging leaves, marigolds, weathered deck planks, cascading water, and distant peaks with temporal lookahead.
  • When to use RazorNext on video: On soft 480p/720p footage packed with fine foliage, blossoms, stone textures, or moving water (VE_CLI.exe upscale input.mp4 --engine razornext --scale 2), RazorNext pulls detail from neighboring frames via its temporal lookahead window, avoiding the frame-to-frame shimmer common in single-frame upscalers and soft Lanczos resizing.
  • When to switch models or pair cleanup: Use PeakMagnetar on clean, low-noise camera clips; pick OptiPulsar when batch speed matters; switch to RTX Video Super Resolution (Quality 3–4) when streaming compression blocks dominate; or place Denoising (ClearMagnetar) above Upscale if the source has visible ISO grain or shadow blocking before scaling.

RTX Video Super Resolution excels on streaming captures, compression artifacts, and soft web video using an NVIDIA RTX GPU (>= 7.5).

Control Options Default Notes
VSR Quality 1 - Low, 2 - Medium, 3 - High, 4 - Ultra 4 - Ultra Higher levels remove stronger compression blocking & ringing
Sizing Mode Scale factor, Custom dimensions Scale factor Output cannot be smaller than the source dimensions
Scale factor 1×, 1.5×, 2×, 3×, 4× 2× 1× enhances clarity & de-blocks without changing resolution
Custom dimensions Width (px) (3840), Height (px) (2160), Lock Aspect Ratio (On) 3840 × 2160, Locked Enter exact pixel targets while maintaining or unlocking aspect

Scaling provides exact mathematical resizing from 25% down to 1600% up without neural hallucination.

Control Image Options Video Options Default
Filter Lanczos4, Area, Bicubic, Bilinear, Nearest Spline36, Lanczos, Bicubic, Area, Bilinear Lanczos4 (Image) / Spline36 (Video)
Scale Source (Original) (100%), 1600%, 800%, 400%, 300%, 200%, 175%, 150%, 125%, 75%, 50%, 25% Same Source (Original)
  • Filter notes: Use Nearest on pixel art to keep razor-sharp square texels; use Area when downscaling (25%–75%) to avoid moiré shimmer; use Lanczos4 or Spline36 for balanced photographic clarity.
  • Minimum output: 64 × 64 px.

Resolution & Codec Ceilings (16,384 px vs. 32,768 px)

Section titled “Resolution & Codec Ceilings (16,384 px vs. 32,768 px)”

Pipeline engines and export formats enforce distinct maximum side lengths—keep these caps in mind before chaining 4×–8× factors onto 4K sources:

Engine / Target Maximum Side Length Key Constraint
DLSS Super Resolution / DLSS Neural Rendering / DLSS Frame Generation / FSR Frame Generation 16,384 px NGX / DirectX 12 texture ceiling
Upscale / RTX Video Super Resolution / Scaling (Pipeline) 32,768 px WebP export caps at 16,383 px per side
NVENC H.265 / AV1 (Video Export) 8,192 × 8,192 px > 8,192 px requires software AV1, VP9, ProRes, or FFV1
H.264 (NVENC / Software) 4,096 × 4,096 px Requires even dimensions; switch to H.265 / AV1 for 4K+