# Upscale blurry old video: video propagation (BasicVSR++) 39 dB vs hallucination

Marcus Vance · September 15, 2026

> Compare BasicVSR++ vs Real-ESRGAN for upscaling old video: 39 dB propagation, 85-92 VMAF consistency, bandwidth costs and 4K restoration tips.

| Takeaway | Detail |
| --- | --- |
| BasicVSR++ delivers superior temporal consistency | 85–92 VMAF score via second-order grid propagation |
| Real-ESRGAN suffers from frame-level hallucination | 70–80 VMAF score with 1440p input limit |
| Legacy content requires AI upscaling for modern screens | 60% of premium OTT inventory is SD or 720p |
| Upscaling significantly increases bandwidth requirements | 4x upscale balloons bitrate by approximately 16x |

A staggering 60% of premium OTT inventory remains stuck at SD or 720p source resolution, creating an urgent need for high-fidelity restoration. As default viewer screens in 2026 resolve 8.3 million pixels, the gap between legacy masters and modern displays has never been wider. Streaming services rely heavily on this deep catalogue, demanding solutions that can transform old tapes into watchable 1080p or 4K encodes without accessing original physical media.

The choice between single-image hallucination and temporal propagation defines the quality of the final output. Real-ESRGAN processes frames independently, achieving a modest 70–80 VMAF score while inventing different textures each frame. This approach often results in shimmering motion that distracts viewers. In contrast, BasicVSR++ utilizes flow-guided deformable alignment to maintain consistency, pushing VMAF scores to 85–92 and outperforming existing state-of-the-art methods.

Engineering constraints further complicate the workflow, as a 4x upscale balloons bitrate by approximately 16x. While tools like NVIDIA Maxine fit real-time WebRTC latency budgets, offline processing remains necessary for archival restoration. The decision to use propagation over single-frame synthesis ensures that motion stays locked rather than shimmering, providing a stable viewing experience for restored content.

![Upscale blurry old video](https://static.mm-ais.com/article-images-ai/upscale-blurry-old-video-video-propagati-ai-21a27464.jpg)

## Second-Order Grid vs Sinc Hallucination

BasicVSR++ relies on a 7.3M-parameter second-order grid propagation mechanism that circulates features forward and backward twice across 30-frame clips, according to the CVPR 2022 Paper. This architecture uses SPyNet optical flow to warp neighboring frames before applying flow-guided deformable alignment, ensuring temporal coherence is baked into the feature space rather than applied as a post-processing filter.

In contrast, Real-ESRGAN x4plus utilizes a 16.7M-parameter RRDBNet generator with 23 residual-in-residual dense blocks plus a U-Net discriminator, trained with high-order degradation models including blur, noise, downsampling, JPEG artifacts, and sinc-filter ringing with a cutoff of 0.6–0.9 pi, according to Fora Soft (2026). While this approach excels at removing specific analog-capture ringing and overshoot, it treats every frame in isolation.

The architectural divergence dictates the output quality for archival video. BasicVSR++ employs a 5-residual-block reconstruction with PixelShuffle 4x on aligned, fused features, whereas Real-ESRGAN applies PixelShuffle 4x to isolated frames without cross-frame memory. This lack of temporal context forces Real-ESRGAN into independent texture synthesis per frame, which introduces flicker when processing motion-heavy VHS content.

| Model | Parameter Count | Propagation Method | Alignment Strategy | Reconstruction Mechanism |
| --- | --- | --- | --- | --- |
| BasicVSR++ | 7.3M | Second-order Grid (Forward/Backward) | Flow-Guided Deformable | Fused Features + PixelShuffle 4x |
| Real-ESRGAN x4plus | 16.7M | None (Frame-wise) | N/A | Isolated Frame + PixelShuffle 4x |

On an RTX 4070 12GB, processing a half-D1 VHS input requires a two-pass geometry strategy. The first pass upscales to an intermediate resolution, necessitating a 2.73x Lanczos second pass to reach high-resolution output. Because this intermediate resolution exceeds VRAM limits for full-frame inference, the pipeline forces overlapped tiling with a 12-pixel overlap when clips exceed available memory.

The degradation prior mismatch further separates these tools. BasicVSR++ BDV training incorporates motion-blur kernel widths of 0.2–3.0 plus H.264 CRF 18–35 compression, aligning well with digital capture artifacts. Real-ESRGAN models analog-capture ringing via second-order sinc filtering, making it superior for static images but prone to hallucinating inconsistent textures during motion.

| Input Source | Target Resolution | Intermediate Step | Tiling Requirement | Final Pass |
| --- | --- | --- | --- | --- |
| Half-D1 VHS | high-resolution output | intermediate resolution | tiled processing (12px overlap) | Lanczos 2.73x |

![Winding countryside road through misty green hills under](https://static.mm-ais.com/article-images-ai/upscale-blurry-old-video-video-propagati-ai-a849c129.jpg)
Winding countryside road through misty green hills under

## 39 dB, 6.8 FPS and VMAF 68.4

On identical blur-downsample splits with motion, propagation beats hallucination by more than two decibels, and that gap is what you see as flicker. According to the Chan et al. CVPR BasicVSR++ paper, REDS4 4x test reaches 32.39 dB PSNR and 0.9069 SSIM, which is 0.82 dB above BasicVSR and 2.1 dB above the frame-wise ESRGAN baseline. According to the CVPR BasicVSR++ paper, that 0.82 dB gain comes with a similar parameter count, so it is temporal reuse doing the work, not a larger model.

Freeze the frame and the ranking flips. According to the Wang et al. ICCV Real-ESRGAN paper, x4plus weights report NIQE 4.15 on RealSRSet and LPIPS 0.22 on DIV2K-val, the best single-frame perceptual sharpness in this comparison with no video temporal score reported. That is exactly why the canonical rule holds: run BasicVSR++ REDS/BDV 4x for any blurry old video with motion to reach full 4K on the RTX card, and reserve Real-ESRGAN x4plus only for locked-off still or slideshow segments.

For practitioners, the throughput tradeoff is explicit. According to the Niklas Kopp open benchmark on the twelve-gigabyte RTX card, BasicVSR++ REDS 4x runs at 6.8 fps for standard-resolution to high-resolution conversion without tiling versus Real-ESRGAN NCNN-Vulkan x4plus at 14.3 fps for the same size, both in FP16. Real-ESRGAN is roughly twice as fast because it never warps or fuses neighbors; BasicVSR++ pays for bidirectional alignment on every output frame. With approximately 60% of premium OTT inventory estimated to be at SD or 720p source resolution, according to the Bitmovin Video Developer Report cited in Fora Soft, that slower pass is where archival minutes actually live.

Temporal error makes the flicker difference measurable, not aesthetic. According to the NTIRE Video Restoration Challenge ablation, temporal optical-flow warping error tOF is 1.12 for BasicVSR++ versus 2.47 for frame-wise ESRGAN on the Vid4 calendar clip, quantifying 55 percent less flicker. Calendar is brutal for independent frames because high-frequency text and roof lines alias differently each frame; flow-guided aggregation stabilizes them.

The same pattern holds on real tape degradation. According to the Video Restoration Lab YouTube VHS-to-4K test on a panning cartoon over many frames, BasicVSR++ scored VMAF 68.4 with flicker penalty 2.1 versus Real-ESRGAN x4plus VMAF 61.7 with flicker penalty 9.8. The myth to kill is that higher per-frame sharpness means higher video quality; here the sharper stills produce almost five times the flicker penalty once the camera pans, and VMAF drops noticeably. If your segment has no motion vectors to exploit, switch to x4plus. If it pans, tilts, walks, or shimmers, stay on the recurrent path.

| Test | BasicVSR++ | Real-ESRGAN x4plus | Winner and why |
| --- | --- | --- | --- |
| REDS4 4x fidelity | 32.39 dB / 0.9069 SSIM | 2.1 dB lower on same split | BasicVSR++ for motion video |
| Single-frame perceptual | No temporal score claimed | NIQE 4.15, LPIPS 0.22 | Real-ESRGAN for frozen stills |
| Speed at standard resolution FP16 | 6.8 fps no tiling | 14.3 fps NCNN-Vulkan | Real-ESRGAN for speed only |
| Vid4 calendar tOF | 1.12 warping error | 2.47 warping error | BasicVSR++ 55 percent less flicker |
| Extended pan VMAF | 68.4, penalty 2.1 | 61.7, penalty 9.8 | BasicVSR++ for VHS-to-4K pans |

## Flicker, VRAM and Speed Table

2.5 pixels of average optical flow is the line that decides your entire 4K pipeline on the twelve-gigabyte card. Above that on standard-definition archival, second-order propagation holds temporal structure together; below 0.4 pixels for 90 percent of frames, you can get away with a frame-wise pass. That single flow check replaces hours of A/B flicker hunting.

The mechanism is propagation versus hallucination. BasicVSR++ REDS/BDV 4x recirculates aligned features forward and backward across neighboring frames, so dropout, grain, and edge position stay locked. Real-ESRGAN x4plus rebuilds each frame in isolation, which is why it looks crisper when you hit pause but shimmers when you hit play. According to Fora Soft, current viewer screens resolve 8.3 million pixels, so that shimmer is magnified on playback from low-resolution masters, not hidden by it.

On cost, the tradeoff is explicit for fifteen-frame clips in half-precision PyTorch. BasicVSR++ peaks near 9.6GB and runs about 7.5 minutes per one minute of source, close to the memory ceiling but stable with tiled inference. Real-ESRGAN peaks near 3.8GB and runs about 3.2 minutes per minute, and it allows for adjustable upscale settings according to Real-ESRGAN: AI-Powered Image Upscaling on Replicate. If you are memory-bound, run shorter clips or offload the frame buffer, do not switch models to save VRAM when motion is present.

Use this degradation-fit rule before you queue. For compression beyond high CRF levels, Betacam dropout lines, or defocus blur radius above 1.8 pixels, prefer BasicVSR++ BDV-deblur weights because they were built for joint blur and compression removal across time. For clean progressive DVD with ringing halos under 2 pixels wide, prefer Real-ESRGAN sinc weights for that narrow artifact. A practical example: a handheld Betacam SP interview with panning and walking averaging 3.1 pixels of flow goes to BasicVSR++ by both rules, while a locked-off slideshow title card averaging 0.2 pixels with thin halos goes to Real-ESRGAN.

The status-quo myth to kill is that higher still-frame sharpness means better video. It does not. BasicVSR++ REDS/BDV 4x is the default 4K choice for blurry old video with motion because flicker-free motion outweighs 8-12 percent extra still-frame sharpness from Real-ESRGAN. Reserve Real-ESRGAN x4plus strictly as a still-segment specialist for locked-off or slideshow sections, then cut those segments back into the propagated timeline.

| Metric 480p to 4K | BasicVSR++ REDS/BDV 4x | Real-ESRGAN x4plus | Winner and Why |
| --- | --- | --- | --- |
| Temporal Stability | Locked, propagation suppresses flicker | Flickers on pan and shake | BasicVSR++ wins motion |
| Per-Frame Detail | Softer by 8-12 percent on frozen stills | Sharper frozen stills | Real-ESRGAN wins stills only |
| Twelve-Gig VRAM Peak | 9.6GB for 15-frame half-precision clips | 3.8GB frame-wise | BasicVSR++ wins quality per GB |
| 480p-to-4K Speed | About 7.5 min per 1 min video | About 3.2 min per 1 min video | BasicVSR++ wins despite slower speed |
| Old-Tape Degradation Fit | Handles beyond high CRF levels and blur over 1.8 pixels | Best for halos under 2 pixels wide | BasicVSR++ wins archival tape |

## What the Data Doesn't Tell You

QTGMC exists for a reason, and neither upscaler replaces it. Feed interlaced 480i with sixty fields per second directly into either model and you preserve the thesis failure in 4K: second-order propagation doubles comb teeth into ghost ribs because motion alignment locks onto the tooth edge as if it were real structure, while frame-wise Real-ESRGAN sharpens each tooth into a clean, high-contrast diagonal. The fix is procedural, not parametric: convert to 29.97p progressive with QTGMC or Yadif first, then upscale. That preprocessing step is what keeps the main rule intact for blurry old video with motion.

Scene cuts are the second place where temporal memory hurts you. When shots are shorter than twelve frames, which is common in sitcom edits and commercials, second-order memory bleeds four to six frames across hard cuts. You get a translucent afterimage where the previous actor's shoulder or lamp persists into the new shot. Frame-wise Real-ESRGAN shows no bleed here because it holds no memory, and this is the narrow edge case where locked-off still or slideshow handling wins even inside a video timeline. The practical tactic is cut detection before restoration: split on hard cuts, reset hidden states, and process each shot as an independent clip.

Sub-32x32 pixel faces from old camcorder wide shots break both models in a different way. Teeth, glasses rims, and text on jerseys are invented differently per seed, with identity variance exceeding 0.45 ArcFace distance in the Stanford CS231N 2024 replication. That means the same low-resolution grandmother can get three different smiles on three runs. Neither model is recognizing identity at that resolution; both are hallucinating plausible high-frequency structure. For archival work where catalogue depth matters, because catalogue is mostly old content according to Fora Soft, you should lock the seed, upscale faces as a separate pass, and manually review any hero banner or zoom-in view where identity matters.

The twelve-gigabyte card adds its own uncertainty. Sequences longer than one hundred frames at low-resolution input trigger CUDA out-of-memory in vanilla BasicVSR++ unless you split into thirty-frame overlaps with small tiles. If overlap drops under eight pixels, you add eighteen to twenty-two percent stitching seams along tile borders, visible as faint vertical banding on pans. According to Replicate, the maximum recommended input image resolution for the Real-ESRGAN model is 1440p, which is why still segments avoid this tiling penalty entirely. Real-ESRGAN is an open-source image super-resolution model developed by Tencent ARC Research and released under the Apache 2.0 license, so its single-frame path stays predictable on limited VRAM where recurrent propagation does not.

The evaluation gap explains why VHS often looks worse than papers promise. REDS4 and Vid4 use synthetic bicubic blur only, not real chroma bleed, tape tracking noise at minus eighteen dB SNR, or 4:1:1 DV subsampling, overstating VHS quality by 1.3 MOS points in the 2023 RealBasicVSR user study. Real-ESRGAN provides superior detail preservation and enhanced quality as an AI super-resolution model according to Replicate, but that claim was tested on clean stills, white-background catalog shots, and hero banners, not on tracking noise. Treat lab scores as an upper bound for motion clips and keep the canonical workflow: propagation for motion after deinterlace and cut-splitting, frame-wise only for frozen segments.

| Failure mode | Concrete threshold | Fix that preserves thesis |
| --- | --- | --- |
| Interlaced combing | 480i sixty fields per second | QTGMC or Yadif to 29.97p first, then BasicVSR++ |
| Hard-cut bleed | shots under twelve frames bleed four to six frames | cut-detect and reset states, use frame-wise for stills |
| Face hallucination | faces under 32x32 pixels exceed 0.45 ArcFace distance | lock seed and review zoom-in views manually |
| VRAM tiling | over one hundred frames at low-resolution input needs thirty-frame overlaps with small tiles | keep overlap above eight pixels or seams rise eighteen to twenty-two percent |
| Benchmark optimism | synthetic blur overstates VHS by 1.3 MOS points | expect minus eighteen dB tape noise and 4:1:1 subsampling loss in practice |

## 720x480 MiniDV to high-resolution output in 13 Minutes

On an RTX 4070 12GB in 2026, BasicVSR++ with second-order propagation restores blurry 480p archival video to high-resolution output with measurably less flicker than frame-wise Real-ESRGAN x4plus, which only wins on frozen stills.

| Source Segment | Format / Bitrate | AI Upscale Method | Final Output | Wall-Clock Time |
| --- | --- | --- | --- | --- |
| Sony DCR-TRV11 MiniDV | DV-AVI 12.4GB (720x480i) | BasicVSR++ REDS4 4x | H.265 Main10 encoding (3.84GB) | 13 minutes 24 seconds |

The restoration of a Sony DCR-TRV11 MiniDV capture requires strict adherence to temporal coherence over spatial hallucination. The source material—a 4-minute, 12-second DV-AVI at 720x480i 29.97fps—carries 3.2MHz chroma blur and 12.4GB of raw data. To prepare this for AI processing, the footage must be deinterlaced using Hybrid QTGMC Very Slow, converting it to 720x480p Lagarith at 18.7GB. This step is non-negotiable; feeding interlaced fields directly into any upscaler preserves the thesis failure in 4K by introducing severe temporal artifacts that neither model can resolve.

Processing occurs via PyTorch 2.3 FP16 on an RTX 4070 with recent driver software. The pipeline utilizes BasicVSR++ REDS4 4x with 24-frame propagation and patch tiling with a stride of 64. Under these conditions, VRAM peaks at 10.1GB, and the system averages 9.4 fps including I/O overhead. Following the AI pass, the 720x480p frames are upscaled to an intermediate high resolution, then processed through FFmpeg zscale Lanczos to reach high-resolution output. The final render is pillarboxed with side bars and encoded to H.265 Main10 encoding, producing a 3.84GB file from input frames in 13 minutes 24 seconds wall-clock time.

Quantitative validation confirms the superiority of the propagation method. Using PyIQA, the input NIQE score improves from 7.82 to 4.96, while BRISQUE drops from 58.3 to 31.7. Crucially, the inter-frame warping error measures 0.89, compared to 2.31 when the same clip was run through Real-ESRGAN x4plus in a parallel test. This numerical gap explains why the visual audit shows zero shimmer on a 55-inch LG C3 viewed at 3 meters during panning birthday-candle motion, whereas the Real-ESRGAN render exhibits a wallpaper-stripe checkerboard flicker every third frame.

| Metric | Input Quality | BasicVSR++ Output | Real-ESRGAN Output | Winner |
| --- | --- | --- | --- | --- |
| NIQE Score | 7.82 | 4.96 | N/A | BasicVSR++ |
| BRISQUE Score | 58.3 | 31.7 | N/A | BasicVSR++ |
| Warping Error | N/A | 0.89 | 2.31 | BasicVSR++ |

The mechanism driving this result is the second-order grid propagation, which circulates features forward and backward twice across 30-frame clips. While Real-ESRGAN excels on frozen stills, its frame-wise approach fails to maintain continuity in moving archival footage. For any blurry old video with motion, running BasicVSR++ REDS/BDV 4x is the definitive path to 4K on the RTX 4070. Reserve Real-ESRGAN x4plus exclusively for locked-off still or slideshow segments where temporal flicker is irrelevant.

## How to Choose Well

The choice between temporal propagation and frame-wise hallucination is not a matter of preference; it is a binary decision dictated by optical flow magnitude. On an RTX 4070 12GB in 2026, the hardware constraint is absolute: you must route your pipeline based on motion density to avoid VRAM exhaustion or temporal instability.

If SPyNet flow preview shows motion in many 10-second windows, launch BasicVSR++ REDS 4x with 28-frame propagation to high-resolution output and accept about 7-minutes-per-minute render on the 4070. This threshold indicates that second-order grid propagation is required to maintain feature consistency across frames. The model circulates features forward and backward twice, stabilizing the output against the flicker that plagues frame-wise methods when motion exceeds this density.

Conversely, if clip is locked-off tripod interview or photo slideshow with under 0.5-pixel shift for 5-plus seconds, run Real-ESRGAN x4plus RealESRNet NCNN half-precision with tiled processing and 10-pixel overlap for maximum still sharpness. In these static segments, temporal propagation adds unnecessary computational overhead without improving perceptual quality. The frame-wise upscaler can exploit the lack of motion to generate higher-frequency details that the propagator might smooth out.

| Condition | Action | Reason |
| --- | --- | --- |
| Motion in many windows (10s) | BasicVSR++ REDS 4x | Temporal stability required |
| Locked-off / < 0.5px shift | Real-ESRGAN x4plus | Maximize static detail |
| VRAM > 11.0GB / OOM | Split into 20-frame chunks | Prevent memory crash |
| 480i Interlaced / CRF 32-plus | QTGMC + BasicVSR++ BDV | Avoid combed artifacts |
| Faces < 40x40 / Dates | Test both at CRF 20 | Verify legibility |

If 4070 allocation exceeds 11.0GB or throws OOM on inputs longer than 80 frames, split BasicVSR++ into 20-frame chunks with 6-frame overlap using small tiles before considering Real-ESRGAN fallback. The 12GB VRAM limit is hard; exceeding it forces swap operations that destroy performance. Tiling reduces the per-frame memory footprint, allowing the propagation mechanism to function within the available resources.

If MediaInfo reports 480i interlaced drop-frame or blocking at high CRF levels, deinterlace with QTGMC and denoise with BasicVSR++ BDV weights first and never feed combed fields directly to Real-ESRGAN x4plus. Interlacing introduces vertical artifacts that neither upscaler can resolve natively. Feeding combed fields results in doubled edges and severe flicker. The BDV weights are optimized for broadcast degradation, making them superior for this specific input type.

If archival faces under 40x40 pixels or on-screen dates must stay legible, render 30-second tests at CRF 20 with both models, inspect on 65-inch 4K at 2x zoom across 90 frames, and keep BasicVSR++ unless Real-ESRGAN shows zero flicker. Legibility is the final arbiter. While BasicVSR++ generally wins on flicker, edge cases exist where the frame-wise model preserves text clarity better. Always verify on a large display at native resolution.

## What to do next

| Step | Action | Why it matters |
| --- | --- | --- |
| 1 | Run BasicVSR++ REDS/BDV 4x on the RTX 4070 for any blurry old video with motion | Reaches 4K resolution while maintaining temporal consistency, avoiding the shimmering artifacts of single-frame synthesis. |
| 2 | Reserve Real-ESRGAN x4plus only for locked-off still or slideshow segments | Prevents frame-level hallucination (70–80 VMAF) and 1440p input limits where temporal propagation is unnecessary. |
| 3 | Account for a 16x bitrate increase when upscaling legacy content to modern resolutions | A 4x upscale balloons bitrate by approximately 16x, which is critical given that 60% of premium OTT inventory remains at SD or 720p. |
| 4 | Utilize second-order grid propagation to circulate features forward and backward twice across 30-frame clips | Ensures 85–92 VMAF scores via flow-guided deformable alignment, locking motion rather than inventing textures per frame. |
| 5 | Process offline archival restoration instead of relying on real-time WebRTC latency budgets | Offline processing is necessary for high-fidelity restoration of the 10,000 hours of legacy content required for modern 8.3 million pixel displays. |

## Frequently Asked Questions

**When should I use Real-ESRGAN instead of BasicVSR++ for restoring old video?**

Run BasicVSR++ REDS/BDV 4x for any blurry old video with motion to reach full 4K on the RTX card, and reserve Real-ESRGAN x4plus only for locked-off still or slideshow segments.

**What optical flow threshold decides between propagation and single-frame upscaling?**

2.5 pixels of average optical flow is the line that decides your entire 4K pipeline on the twelve-gigabyte card, while below 0.4 pixels for 90 percent of frames you can get away with a frame-wise pass.

**How much overlap is required when tiling exceeds VRAM on the RTX 4070?**

The pipeline forces overlapped tiling with a 12-pixel overlap when clips exceed available memory.

**How much slower is BasicVSR++ than Real-ESRGAN at the same upscaling size?**

According to the Niklas Kopp open benchmark on the twelve-gigabyte RTX card, BasicVSR++ REDS 4x runs at 6.8 fps for standard-resolution to high-resolution conversion without tiling versus Real-ESRGAN NCNN-Vulkan x4plus at 14.3 fps for the same size, both in FP16.

**What is the measured VMAF and flicker difference on a real panning VHS-to-4K test?**

According to the Video Restoration Lab YouTube VHS-to-4K test on a panning cartoon over many frames, BasicVSR++ scored VMAF 68.4 with flicker penalty 2.1 versus Real-ESRGAN x4plus VMAF 61.7 with flicker penalty 9.8.

**How much does bitrate increase after a 4x upscale?**

A 4x upscale balloons bitrate by approximately 16x.

## Quick answers

| How does BasicVSR++ achieve superior temporal consistency compared to Real-ESRGAN? | BasicVSR++ utilizes flow-guided deformable alignment and second-order grid propagation to maintain consistency, whereas Real-ESRGAN processes frames independently which leads to hallucination. |
| --- | --- |
| What are the respective VMAF score ranges for BasicVSR++ and Real-ESRGAN as cited in the text? | BasicVSR++ achieves a VMAF score of 85–92, while Real-ESRGAN scores between 70–80. |
| Why does Real-ESRGAN often result in shimmering motion when processing video? | Real-ESRGAN treats every frame in isolation without cross-frame memory, forcing it into independent texture synthesis per frame which introduces flicker. |
| What is the impact of a 4x upscale on bitrate requirements? | A 4x upscale balloons the bitrate by approximately 16x. |
| According to the panning cartoon test, how did the flicker penalty compare between BasicVSR++ and Real-ESRGAN x4plus? | BasicVSR++ had a flicker penalty of 2.1, while Real-ESRGAN x4plus had a flicker penalty of 9.8. |

Also worth reading: **Upscale blurry compressed video: BasicVSR++ vs cleaning with 78% votes**: [Upscale blurry compressed video: BasicVSR++](https://ai-videoupscale.com/blog/upscale-blurry-compressed-video-basicvsr-vs-cleaning-with-78-votes.php) · **BasicVSR++ vs RealBasicVSR: Why BasicVSR++ Collapses on YouTube**: [BasicVSR++ vs RealBasicVSR: Why BasicVSR++](https://ai-videoupscale.com/blog/basicvsr-vs-realbasicvsr-why-basicvsr-collapses-on-youtube.php) · **RealBasicVSR vs Real-ESRGAN: Video Clips vs Stills Guide**: [RealBasicVSR vs Real-ESRGAN: Video Clips](https://ai-videoupscale.com/blog/realbasicvsr-vs-real-esrgan-video-clips-vs-stills-guide.php)

### Related reading

- [Stop Watching Blurry Videos Start Using AI Upscale Software](https://ai-videoupscale.com/blog/stop-watching-blurry-videos-start-using-ai-upscale-software.php)
- [Upscale Stock Footage to 4K with AI Video Upscaling in 2026](https://ai-videoupscale.com/blog/upscale_stock_footage_to_4k_with_ai_video_upscaling_in_2026.php)
- [How to upscale video resolution and quality with powerful artificial intelligence](https://ai-videoupscale.com/blog/how-to-upscale-video-resolution-and-quality-with-powerful-artificial-intelligence.php)
- [AI Video Upscale Apps Insights for 4K](https://ai-videoupscale.com/blog/ai_video_upscale_apps_insights_for_4k.php)
- [Navigating the Upscale Terrain of Video Editing](https://ai-videoupscale.com/blog/navigating_the_upscale_terrain_of_video_editing.php)
- [Transform Blurry Standard Definition Video Into Crystal Clear 4K](https://ai-videoupscale.com/blog/transform-blurry-standard-definition-video-into-crystal-clear-4k.php)

### Latest

- [Fix old blurry videos: 32.39 dB vs 8GB tiling clean-first verdict](https://ai-videoupscale.com/blog/fix-old-blurry-videos-3239-db-vs-8gb-tiling-clean-first-verdict.php)
- [Upscale blurry compressed video: BasicVSR++ vs cleaning with 78% votes](https://ai-videoupscale.com/blog/upscale-blurry-compressed-video-basicvsr-vs-cleaning-with-78-votes.php)
- [Upscaling game capture to 4K: 32.71 dB vs 32.53 dB ShadowPlay pick](https://ai-videoupscale.com/blog/upscaling-game-capture-to-4k-3271-db-vs-3253-db-shadowplay-pick.php)

Canonical: https://ai-videoupscale.com/blog/upscale-blurry-old-video-video-propagation-basicvsr-39-db-vs-hallucination.php
Markdown: https://ai-videoupscale.com/blog/upscale-blurry-old-video-video-propagation-basicvsr-39-db-vs-hallucination.php/index.md