Strict rendering measurements performed on physical hardware testbeds across four generations: RTX 2080, RTX 3050, RTX 4070 Ti, and RTX 5080.
Charcoal Macro Photo (4.0x / PSNR 39.40dB), Video (Seagull), Cascade Upscaling, and Slow-Motion comparisons are published on the showcase page.
Removes compression artifacts and restores sharp edges without changing video resolution
Upscales source video by 2.0x (4x pixels; e.g., 3840×2160 from 1080p source)
| GPU / Architecture | Benchmark Test Case | Edition | Speed (fps) | Time (Speedup) | Peak VRAM |
|---|---|---|---|---|---|
| RTX 5080 Blackwell (sm_120 / 16GB) |
Video 1.0x Native Cleansing | PRO / ULTRA | 119.7 fps ➔ 191.6 fps (+60%) | 5.0x Realtime ➔ 8.0x Realtime (20.9s) | +2.34 GB |
| Video 2.0x Super-Resolution (4x Pixels) | PRO / ULTRA | 63.8 fps ➔ 72.0 fps | 2.7x Realtime ➔ 3.0x Realtime (55.8s) | +3.18 GB | |
| ★ Video 4.0x Super-Resolution (16x Pixels) | PRO / ULTRA | 19.5 fps ➔ 25.5 fps (+31%) | 0.8x Realtime ➔ 1.1x Realtime (Faster Than Realtime) | +5.05 GB | |
| Video 2.0x + 2.0x Interpolation (*RTX 40/50 Only) | PRO / ULTRA | 63.0 fps ➔ 75.3 fps | 1.3x Realtime ➔ 1.6x Realtime (106.7s) | +2.30 GB | |
| ★ Video 4.0x Frame Interpolation (*RTX 40/50 Only) | PRO / ULTRA | 124.5 fps ➔ 177.8 fps (+43%) | 1.3x Realtime ➔ 1.9x Realtime (90.4s) | +3.18 GB | |
| ★ Video 2.0x × 4.0x Interpolation (*RTX 40/50 Only) | PRO / ULTRA | 62.8 fps ➔ 77.5 fps (+23%) | 0.7x Realtime ➔ 0.8x Realtime (207.5s) | +1.64 GB | |
| Still Image 2.0x Super-Resolution | PRO / ULTRA | — | 17.59s ➔ 2.25s | 0.85 GB | |
| Still Image 4.0x Super-Resolution | PRO / ULTRA | — | 2.54s ➔ 2.52s | 0.85 GB | |
| ★ Still Image 8.0x Cascade (~67 MP) | PRO / ULTRA | — | 5.87s ➔ 5.85s | 0.85 GB | |
| RTX 4070 Ti Ada Lovelace (sm_89 / 12GB) |
Video 1.0x Native Cleansing | PRO / ULTRA | 79.7 fps ➔ 115.7 fps (+45%) | 3.3x Realtime ➔ 4.8x Realtime (34.7s) | +2.15 GB |
| Video 2.0x Super-Resolution (4x Pixels) | PRO / ULTRA | 36.7 fps ➔ 62.7 fps (+71%) | 1.5x Realtime ➔ 2.6x Realtime (64.0s) | +3.17 GB | |
| ★ Video 4.0x Super-Resolution (16x Pixels) | PRO / ULTRA | 12.4 fps ➔ 19.4 fps (+56%) | 0.5x Realtime ➔ 0.8x Realtime (207.4s) | +4.81 GB | |
| Video 2.0x + 2.0x Interpolation (*RTX 40/50 Only) | PRO / ULTRA | 40.8 fps ➔ 68.5 fps (+68%) | 0.8x Realtime ➔ 1.4x Realtime (117.4s) | +2.35 GB | |
| ★ Video 4.0x Frame Interpolation (*RTX 40/50 Only) | PRO / ULTRA | 87.0 fps ➔ 118.0 fps (+36%) | 0.9x Realtime ➔ 1.2x Realtime (136.2s) | +3.64 GB | |
| ★ Video 2.0x × 4.0x Interpolation (*RTX 40/50 Only) | PRO / ULTRA | 40.0 fps ➔ 71.2 fps (+78%) | 0.4x Realtime ➔ 0.7x Realtime (225.7s) | +1.46 GB | |
| Still Image 2.0x Super-Resolution | PRO / ULTRA | — | 4.84s ➔ 4.61s | 0.85 GB | |
| Still Image 4.0x Super-Resolution | PRO / ULTRA | — | 5.22s ➔ 5.31s | 0.85 GB | |
| ★ Still Image 8.0x Cascade (~67 MP) | PRO / ULTRA | — | 11.65s ➔ 11.12s | 0.85 GB | |
| RTX 2080 Turing (sm_75 / 8GB) |
Video 1.0x Native Cleansing | PRO / ULTRA | 49.0 fps ➔ 57.1 fps (+17%) | 2.0x Realtime ➔ 2.4x Realtime (70.4s) | +1.29 GB |
| Video 2.0x Super-Resolution (4x Pixels) | PRO / ULTRA | 29.6 fps ➔ 46.5 fps (+57%) | 1.2x Realtime ➔ 1.9x Realtime (86.3s) | +1.99 GB | |
| ★ Video 4.0x Super-Resolution (16x Pixels) | PRO / ULTRA | 10.9 fps ➔ 14.4 fps (+32%) | 0.5x Realtime ➔ 0.6x Realtime (278.7s) | +4.69 GB | |
| Still Image 2.0x Super-Resolution | PRO / ULTRA | — | 9.83s ➔ 9.19s | 0.85 GB | |
| Still Image 4.0x Super-Resolution | PRO / ULTRA | — | 7.75s ➔ 8.94s | 0.85 GB | |
| ★ Still Image 8.0x Cascade (~67 MP) | PRO / ULTRA | — | 15.69s ➔ 15.87s | 0.85 GB | |
| RTX 3050 Ampere (sm_86 / 8GB) |
Video 1.0x Native Cleansing | PRO / ULTRA | 25.8 fps ➔ 35.1 fps (+36%) | 1.1x Realtime ➔ 1.5x Realtime (114.3s) | +1.11 GB |
| Video 2.0x Super-Resolution (4x Pixels) | PRO / ULTRA | 11.8 fps ➔ 17.7 fps (+50%) | 0.5x Realtime ➔ 0.7x Realtime (227.5s) | +1.81 GB | |
| ★ Video 4.0x Super-Resolution (16x Pixels) | PRO / ULTRA | 3.9 fps ➔ 5.6 fps (+44%) | 0.2x Realtime ➔ 0.2x Realtime (712.3s Completed) | +4.53 GB | |
| Still Image 2.0x Super-Resolution | PRO / ULTRA | — | 13.39s ➔ 13.27s | 0.85 GB | |
| Still Image 4.0x Super-Resolution | PRO / ULTRA | — | 13.81s ➔ 13.37s | 0.85 GB | |
| ★ Still Image 8.0x Cascade (~67 MP) | PRO / ULTRA | — | 28.76s ➔ 28.99s | 0.85 GB |
Empirical measurements covering 4.0x upscaling, 4.0x frame interpolation (*RTX 40/50 Only), and 8.0x cascade scaling.
16x pixel scaling (4.0x width & height). Dynamic tiling minimizes memory exhaustion risks even on an 8GB RTX 3050 in our test runs.
Synthesizes 3 intermediate motion-compensated optical flow frames per input frame for 4.0x frame-rate output (*RTX 40/50 series only).
Two-stage separated pipeline combining 2.0x upscaling and 4.0x frame interpolation (*RTX 40/50 series only) without VRAM overflow.
Chained 4.0x followed by 2.0x super-resolution. Prevents haloing and outputs ~67 MP resolution tailored for large-format poster printing.
To address common concerns about AI upscaling shimmer or blended interpolation artifacts, we conducted an objective quantitative audit using mathematical signal-processing metrics (PSNR, MAE, Temporal Laplacian, and Motion-Compensated Correlation).
| Asset Tested | PSNR (Signal-to-Noise) | MAE (Mean Absolute Error) | Engineering Interpretation & Status |
|---|---|---|---|
| Still Image 16K Cascade (130 Megapixels) | 39.36 dB | 2.01 | High structural fidelity with minimal synthetic hallucination or structural corruption. |
| Video 4K Super-Resolution (at 8.0s timestamp) | 33.02 dB | 4.61 | Balanced restoration with cleansing, preserving contour sharpness and natural noise reduction. |
| Video 4K Super-Resolution (at 3.0s timestamp) | 31.81 dB | 5.12 | Fine lineart in hair, eyelashes, and pupil highlights preserved as continuous, clean strokes. |
| High-Res Video Crop | 33.47 dB | 4.43 | Clean 100% 8K crop with minimal mosquito noise or flat-field banding. |
| VerificationGPU | h / | VRAM | VRAMSystem Memory | Verdict |
|---|---|---|---|---|
| RTX 5080 (Blackwell 16GB) | 2.95 s | 0.85 GB | 0.00 GB | Complete () |
| RTX 4070 Ti (Ada 12GB) | 5.08 s | 0.85 GB | 0.00 GB | Complete () |
| RTX 2080 (Turing 8GB) | 7.71 s | 0.85 GB | 0.00 GB | Complete () |
| RTX 3050 (Ampere 8GB) | 13.90 s | 0.14 GB | 0.00 GB | Complete () |
Inspect native 1:1 keyframe crops from actual 4K and 8K outputs to verify contour sharpness and compression noise suppression.
Native 1:1 crop comparing 1080p source against Tensor Core 4K/8K super-resolution output. Switch between Slider and A/B Flicker to inspect exact pixel alignment.
Removes compression mosquito noise and haze from web/social-media video at native 1.0x resolution without inflating file dimensions.
Executes 4.0x super-resolution followed by a 2.0x multi-pass Refine stage, preserving fine pencil lines and iris details without bilinear blur or double-edge ringing.
Cascade 24fps video through 4x (96fps) ➔ 4x (384fps) optical-flow interpolation, then conform to 1/16x speed (6.25%) in DaVinci Resolve or Premiere Pro for smooth 24fps cinematic super slow-motion.
This is Photo Edition's signature Supersampling Anti-Aliasing (SSAA) workflow: upscale to 4.0x (8K) or 8.0x (16K / ~130 MP) to reconstruct fine lines and suppress compression noise, then downscale in Photoshop or Clip Studio Paint to your target delivery size (4K, HD, or A4 print).
Traditional AI upscalers suffer from multi-stage bottlenecks: unpacking frames to system RAM ➔ dumping temporary images to disk ➔ feeding the neural network ➔ re-reading ➔ encoding.
Our VRAM Pipeline keeps frame buffers inside GPU VRAM via parallel streaming, minimizing intermediate disk I/O and CPU overhead to reach 194.9 FPS (~8.1x faster than 24fps playback) at 1.0x cleansing on our RTX 5080 testbed (actual throughput varies by GPU).