Hardware Verification Report

Hardware Benchmark Verification Report

Strict rendering measurements performed on physical hardware testbeds across four generations: RTX 2080, RTX 3050, RTX 4070 Ti, and RTX 5080.

📸
Looking for Empirical Quality Showcases & Objective AI Verification Reports?

Charcoal Macro Photo (4.0x / PSNR 39.40dB), Video (Seagull), Cascade Upscaling, and Slow-Motion comparisons are published on the showcase page.

View Empirical Showcase →

Hardware Verification Testbed

Validated GPUs: NVIDIA GeForce RTX 2080 / RTX 4070 Ti / RTX 5080
Onboard VRAM: 8.0 GB / 12.0 GB / 16.0 GB GDDR7
Architectures: Turing / Ada Lovelace / Blackwell
Inference Engine: Proprietary RTX Tensor Core AI Engine (v2.4)
Peak Throughput (RTX 5080): 8.1x Real-Time Video Speed (194.9 fps)
Measured VRAM Delta: +1.1GB to +2.3GB (Designed to Prevent OOM even on 8GB GPUs)

Test 1: 1.0x Fine Cleansing (SSAA Deblur)

Removes compression artifacts and restores sharp edges without changing video resolution

Peak 194.9 FPS
RTX 5080 (Ultra: DirectStream) 194.9 FPS (8.1x Realtime)
194.9 fps
RTX 4070 Ti (Ultra: DirectStream) 128.6 FPS (5.4x Realtime)
128.6 fps
RTX 2080 (Ultra: DirectStream) 57.4 FPS (2.4x Realtime)
57.4 fps

Test 2: Video 2.0x AI Super Resolution (2x Scale / 4x Pixels)

Upscales source video by 2.0x (4x pixels; e.g., 3840×2160 from 1080p source)

Peak 72.7 FPS
RTX 5080 (Ultra: 3840x2160) 72.7 FPS (3.0x Realtime)
72.7 fps
RTX 4070 Ti (Ultra: 3840x2160) 66.0 FPS (2.8x Realtime)
66.0 fps
RTX 2080 (Ultra: 3840x2160) 45.7 FPS (1.9x Realtime)
45.7 fps

Comprehensive 4-Generation GPU Benchmark Matrix

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

🔥 High-Load Benchmark Verification

Empirical measurements covering 4.0x upscaling, 4.0x frame interpolation (*RTX 40/50 Only), and 8.0x cascade scaling.

High Resolution FHD ➔ 7680×4320

Video 4.0x AI Super-Resolution

16x pixel scaling (4.0x width & height). Dynamic tiling minimizes memory exhaustion risks even on an 8GB RTX 3050 in our test runs.

⚡ RTX 5080: 28.5 fps (4.0x output at 1.2x Realtime)
⚡ RTX 4070 Ti: 22.0 fps / RTX 3050: 9.4 fps
🧠 VRAM Delta: ~+2.0–2.3 GB (8GB Safe)
AI Frame Interpolation 4.0x Interpolation (*RTX 40/50 Only)

Video 4.0x AI Frame Interpolation (*RTX 40/50 Only)

Synthesizes 3 intermediate motion-compensated optical flow frames per input frame for 4.0x frame-rate output (*RTX 40/50 series only).

⚡ RTX 5080: 101.9 fps (Over 100 FPS generation)
⚡ RTX 4070 Ti: 82.4 fps
🎯 Supported: RTX 40 / 50 (Maxine VFG)
Concurrent Execution 2.0x Upscale + 4.0x Interpolation (*RTX 40/50 Only)

Video 2.0x Upscale + Frame Interpolation

Two-stage separated pipeline combining 2.0x upscaling and 4.0x frame interpolation (*RTX 40/50 series only) without VRAM overflow.

⚡ RTX 5080: 69.9 fps (4K 96fps processed at ~3x Realtime)
⚡ RTX 4070 Ti: 52.8 fps
🛡️ Stability Verified: Smooth Completion in Test Runs
64x Pixel Scale 1M ➔ 67.6M Pixels

Still Image 8.0x Multi-Pass Super-Resolution

Chained 4.0x followed by 2.0x super-resolution. Prevents haloing and outputs ~67 MP resolution tailored for large-format poster printing.

⚡ RTX 5080: 3.85 s (Inference: 0.18s)
⚡ RTX 4070 Ti: 5.12 s / RTX 3050: 8.40 s
🖨️ Use Case: B0 Large-Format Posters / Fine-Art DTP
Independent Technical Audit

🔬 Objective AI Scientific Audit: Visual Fidelity & Temporal Stability

🔬 Objective Audit Verified

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).

1. Fidelity (Downscaled After vs. Original Before Comparison)

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.
⚡ Temporal Stability (Shimmering & Flicker Suppression)
  • Temporal Laplacian (Flicker Metric): Before 8.67 ➔ After 8.78 (Nearly identical; no significant processing-induced noise spikes)
  • Luminance Flicker Residual: Before 0.481 ➔ After 0.463 (Slightly reduced; steadier than source)
  • Inter-Frame Diff MAE: Before 5.18 ➔ After 5.22 (Preserves source motion dynamics)
  • Conclusion: Confirmed no noticeable line jitter, ghosting, or structural distortion across frames.
🎯 48fps Frame Interpolation 'True Motion Compensation' Verification (Correlation −0.76)
  • Refuting Linear Blending (Crossfading) Fallacies: Measured −0.76 correlation between motion magnitude and deviation from linear frame averaging, confirming true motion-compensated optical flow synthesis rather than cross-fade blending.
  • Inter-Frame Uniformity: Even➔Odd step diff 2.40 / Odd➔Even 2.32 (Symmetric, smooth cadence).
  • No Duplicate Frames: Minimum step diff 0.48 (newly synthesized intermediate frames even in low-motion scenes).

2. Continuous Endurance (10-Run Stability Verification)

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 ()
Verified

📸 Verified Frame Comparison

Real Hardware Output / Native 1:1 Crop

Inspect native 1:1 keyframe crops from actual 4K and 8K outputs to verify contour sharpness and compression noise suppression.

▶
Watch the Full 4K / 48fps Rendered Video on YouTube
Cut clips generated at 0.8MP / 24fps with MiniMax H3 were edited together and exported at 1920×1080 / 24fps in Filmora, then upscaled to 4K and interpolated to 48fps with this tool.
Watch Full Video on YouTube →
Verified 01 / Video Super-Resolution

🎬 Video 2.0x (4K) & 4.0x (8K) AI Super-Resolution Comparison

⚡ RTX 5080: 101.9 fps (4K) 🔬 PSNR 39.36 dB Measured

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.

drag toSuper-Res
Before
After
↔
🔴 BEFORE: 1080p
🟢 AFTER: 4K / 8K Super-Resolution
Verified 02 / Native 1.0x Cleansing

🧹 Video 1.0x SSAA (Native-Resolution Denoise & Sharpen)

⚡ 194.9 fps (~8.1x Realtime) 💾 Native 1.0x

Removes compression mosquito noise and haze from web/social-media video at native 1.0x resolution without inflating file dimensions.

Drag slider to compare
Before
After
↔
🔴 BEFORE: Source Noise
🟢 AFTER: SSAA 1.0x Cleansed
Verified 03 / Large Poster Print Upscaling

🖼️ Still Image 8.0x Multi-Pass Cascade (1376×768 ➔ 11008×6144)

~67 Megapixels Zero Tile Seams

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.

drag toRendering
Before
After
↔
🔴 BEFORE: Source (1376×768)
🟢 AFTER: 8.0x Multi-Pass Cascade
💡 Master Physical RigVerified

🎬 24fps ➔ 384fpsmulti-pass chained ➔ 16cinematicSuper Slow-Motion

24fps ➔ 96fps ➔ 384fps DaVinci / Premiere

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.

Actual Output: 1/16x Super Slow-Motion Playback (24fps ➔ 384fps ➔ 24fps)
🛠️ 3-Step Super Slow-Motion Workflow
1
Pass 1: 24fps ➔ 96fps Interpolation
SuperResStudio (Ultra) performs 4x video frame interpolation.
2
Pass 2: 96fps ➔ 384fps Cascade
Run the 96fps intermediate clip through another 4x interpolation pass to generate a dense 384fps master.
3
Slow Down to 1/16x in NLE (DaVinci Resolve / Premiere)
Drop the 384fps clip onto a 24fps/30fps timeline and set playback speed to 6.25% for fluid 1/16x slow motion.
💡 Tips (Still Image) SSAA

🖼️ Still Image Pro Tip: Supersampling (SSAA) via 8K/16K Downscaling

Up to 16K Crisp Detail

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).

☕
Practical Note on 16K Files:
While our dynamic chunking keeps GPU VRAM usage within safe limits up to 16K (~130 MP), opening 16K images in image editors can consume significant system RAM. For everyday workflows, 4.0x (8K / ~33 MP) ➔ downscale offers the best balance of speed and crispness.
💡
Pro-Tier Technical Q&A

Technical Q&A

Q1: Does 8K (7680×4320) video upscaling risk running out of VRAM (OOM crash)?
A: Our dynamic tiling architecture significantly reduces out-of-memory crash risks (actual stability may vary by driver and system environment). When output resolution exceeds 4K, Automatic GPU Tiling splits each frame into safe rectangular blocks inside GPU VRAM, processes them sequentially on Tensor Cores, and merges them cleanly—keeping VRAM delta around +2GB to +4.5GB even on 8GB GPUs like RTX 3050/2080 (exact behavior depends on your PC environment; please verify with the Free Trial).
Q2: When is 4x Frame Interpolation (24fps ➔ 96fps / 30fps ➔ 120fps) most useful?
A: Ideal for AI-generated video (Kling, Hailuo, Sora) and 24fps anime/slow-motion edits. AI-generated clips often look choppy at 24fps. Running 4x optical-flow interpolation generates 3 intermediate frames per source frame, enabling smooth 96fps/120fps playback or clean slow-motion editing in your NLE.
Q3: Does 8x cascade upscaling (4x then 2x) on still images cause artifacts?
A: Multi-pass cascade Refine suppresses ringing and keeps edges clean. Running Pass 1 (4.0x) followed by Pass 2 (2.0x SSAA Refine) cleanly scales a 1376×768 image up to ~67 MP (11008×6144) for large poster prints without tile seams.
Q4: How stable is VRAM usage during long videos or high-resolution upscaling?
A: Two-stage pipeline isolation and dynamic chunking keep peak VRAM low and stable. By separating super-resolution and frame interpolation stages and releasing GPU memory between chunks, the studio prevents VRAM exhaustion even on demanding high-resolution or long video workloads.
Q5: I recently picked up an entry-level GPU like the RTX 3050 — will it run at practical speeds?
A: Yes, it runs stably. On an RTX 3050 (Ampere sm_86 / 8 GB VRAM), 1.0x video cleansing reaches 42.8 fps (~1.8x realtime) and 2.0x 4K video super-resolution runs at 24.5 fps (~1.0x realtime). (*Optical-flow Frame Interpolation requires RTX 40/50 Series.)
Q6: [Pro Tip] Why upscale to 8K/16K if I only need HD or 4K?
A: Supersampling (SSAA) produces noticeably cleaner lines and edges than native 1:1 processing. Reconstructing fine detail at 4.0x (8K/16K) and then downscaling (Lanczos/Bicubic) to HD or 4K suppresses JPEG compression artifacts and jagged edges.
Q7: Can it handle legacy DVD/VHS/camcorder files (.mpg, .wmv, .vob) and older images (.tif, .bmp)?
A: Yes, decodable legacy formats can be upscaled and modernized to MP4/PNG. Legacy video formats (.mpg, .mpeg, .wmv, .vob, .m2ts, .flv, .avi) and image formats (.tif, .bmp, .tga) supported by the bundled FFmpeg/Pillow can be upscaled, deinterlaced, and exported as standard MP4 (H.264/AAC) or PNG. Please test your specific files with the Free Trial beforehand.

Why Is It So Fast?

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).