Compression Quality Test - Bitrate Recommendation

⚠️

Video Players May Not Work

Video files are not included in this repository (files >50MB cannot be pushed to GitHub). The video players will only work on local machines where the test videos have been generated.
To generate test videos locally: Run ./test-double-compress.sh and ./test-bitrate-floor.sh

Quality Metrics (vs Original 1080p Source)

Excellent: VMAF 93+ / PSNR 38+

Good: VMAF 80-93 / PSNR 35-38

Fair: VMAF 70-80 / PSNR 33-35

Poor: VMAF 60-70 / PSNR 30-33

Bad: VMAF <60 / PSNR <30

PSNR (Peak Signal-to-Noise Ratio): Measures pixel-level fidelity between original and compressed video. Higher values = less distortion. Fast to compute but doesn't model human perception well.
VMAF (Video Multi-Method Assessment Fusion): Netflix's perceptual quality metric that predicts how humans perceive video quality. More realistic than PSNR for real-world quality assessment.

Goal: Find Optimal Bitrate for Video Streaming

The Question

What bitrate cap gives us acceptable quality while minimizing file size and bandwidth?

The Method

Test various bitrates at 720p resolution and measure quality using VMAF (perceptual) and PSNR (pixel-level) metrics.

The Answer

720p @ 2000 kbps - VMAF 71.50 (Fair quality), 7.2 MB per 30s video, ~240 KB/s bandwidth.

Compression Strategy: Client uploads at 720p. Server conditionally re-encodes only when needed (bitrate too high, wrong codec, missing faststart). This avoids unnecessary double compression.

720p: Single vs Double Compression

PSNR measured against original 1080p source. Double compression loses ~0.7 dB.

Original 1080p (Reference)

Resolution1920x1080

Bitrate4,190 kbps

File Size15.0 MB

PSNRReference

Single 720p (CRF 23)

Bitrate3,011 kbps

File Size10.8 MB

PSNR vs Source34.99 dB (Fair)

Double 720p (CRF 23 x2)
                    
                Bitrate2,634 kbps
File Size9.4 MB
PSNR vs Source34.25 dB (Fair)
PSNR Loss-0.74 dB

480p: Single vs Double Compression

Resolution downscale contributes more to quality loss than double compression.

Original 1080p (Reference)

Resolution1920x1080

Bitrate4,190 kbps

PSNRReference

Single 480p

Bitrate1,596 kbps

File Size5.7 MB

PSNR vs Source32.09 dB (Poor)

Double 480p

Bitrate1,307 kbps

File Size4.7 MB

PSNR vs Source31.40 dB (Poor)

PSNR Loss-0.70 dB

360p: Single vs Double Compression

Original 1080p (Reference)

Resolution1920x1080

Bitrate4,190 kbps

PSNRReference

Single 360p

Bitrate1,003 kbps

File Size3.6 MB

PSNR vs Source30.64 dB (Poor)

Double 360p

Bitrate877 kbps

File Size3.1 MB

PSNR vs Source30.17 dB (Poor)

PSNR Loss-0.47 dB

320p: Single vs Double Compression

At 320p, quality is primarily limited by resolution, not compression method.

Original 1080p (Reference)

Resolution1920x1080

Bitrate4,190 kbps

PSNRReference

Single 320p

Bitrate958 kbps

File Size3.4 MB

PSNR vs Source30.09 dB (Poor)

Double 320p

Bitrate716 kbps

File Size2.6 MB

PSNR vs Source29.71 dB (Bad)

PSNR Loss-0.38 dB

240p: Single vs Double Compression

Original 1080p (Reference)

Resolution1920x1080

Bitrate4,190 kbps

PSNRReference

Single 240p

Bitrate566 kbps

File Size2.0 MB

PSNR vs Source29.05 dB (Bad)

Double 240p

Bitrate453 kbps

File Size1.6 MB

PSNR vs Source28.76 dB (Bad)

PSNR Loss-0.29 dB

VMAF Perceptual Quality Results (Human Perception)

VMAF (Video Multi-Method Assessment Fusion) - Netflix's perceptual quality metric that predicts how humans perceive video quality. More realistic than PSNR.

Test	Bitrate	VMAF	Quality	Note
720p Single (baseline)	3,004 kbps	78.26	Fair	Best possible
720p Double (CRF)	2,628 kbps	73.55	Fair	-4.7 from single
720p @ 2500k	2,618 kbps	73.49	Fair	Same as CRF
720p @ 2000k ⭐	2,219 kbps	71.50	Fair	Best perceptual
720p @ 1500k	1,667 kbps	67.98	Poor	Mobile acceptable
720p @ 1000k	1,111 kbps	61.72	Poor	TikTok level
720p @ 800k	887 kbps	57.43	Bad	Below acceptable

PSNR vs VMAF Comparison:
• PSNR said 1500k = "Good" (35 dB) but VMAF says "Poor" (68)
• VMAF is stricter because it models human perception
• For perceptual "Fair" quality (VMAF 70+), use 2000 kbps

720p Bitrate Cap Recommendation

Updated Recommendation: 2000 kbps - Based on VMAF perceptual quality for "Fair" grade (VMAF 70+).

Bitrate Cap	PSNR	VMAF	Quality	File Size (30s)	Use Case
No cap (CRF 23)	34.25 dB	73.55	Fair	9.4 MB	Maximum quality
2000 kbps ⭐	36.22 dB	71.50	Fair (VMAF 70+)	7.2 MB	Recommended
1500 kbps	35.25 dB	67.98	Poor (VMAF)	5.4 MB	Size priority
1000 kbps	33.89 dB	61.72	Poor	3.5 MB	TikTok parity
800 kbps	33.13 dB	57.43	Bad	2.8 MB	Not recommended

⭐ RECOMMENDED

2000 kbps

VMAF 71.50 (Fair)

Size Priority

1500 kbps

VMAF 67.98 (Poor)

TikTok Level

1000 kbps

VMAF 61.72 (Poor)

Why 2000 kbps?
• VMAF 71.50 = "Fair" perceptual quality (mobile/social media level)
• PSNR 36.22 dB = "Good" pixel-level fidelity
• 23% smaller than no-cap CRF 23 (7.2 MB vs 9.4 MB for 30s)
• ~240 KB/s bandwidth requirement

Bitrate Floor Test (720p Double Compressed)

All videos: 1080p source → 720p client compress → backend bitrate-capped + faststart

2000 kbps

PSNR36.22 dB

Size7.2 MB

1500 kbps

PSNR35.25 dB

Size5.4 MB

1000 kbps (TikTok)
                    
                PSNR33.89 dB
Size3.5 MB

800 kbps (Min TikTok)
                    
                PSNR33.13 dB
Size2.8 MB

600 kbps

PSNR32.16 dB

Size2.0 MB

300 kbps

PSNR29.98 dB

Size1.0 MB

Summary: Absolute Quality vs Original 1080p

Resolution	Type	Bitrate	Size (30s)	PSNR vs Source	Grade
1080p	Original	4,190 kbps	15.0 MB	Reference	-
720p	Single	3,011 kbps	10.8 MB	34.99 dB	Fair
720p	Double	2,634 kbps	9.4 MB	34.25 dB	Fair
480p	Single	1,596 kbps	5.7 MB	32.09 dB	Poor
480p	Double	1,307 kbps	4.7 MB	31.40 dB	Poor
360p	Single	1,003 kbps	3.6 MB	30.64 dB	Poor
360p	Double	877 kbps	3.1 MB	30.17 dB	Poor
320p	Single	958 kbps	3.4 MB	30.09 dB	Poor
320p	Double	716 kbps	2.6 MB	29.71 dB	Bad
240p	Single	566 kbps	2.0 MB	29.05 dB	Bad
240p	Double	453 kbps	1.6 MB	28.76 dB	Bad

Key Finding: Double compression adds only ~0.5-0.7 dB loss on top of resolution downscaling. The major quality loss comes from resolution reduction, not the double compression itself.