Research · Results
Saliency-guided perceptual HEVC coding
x265 quantOffsets · TASED-Net / ViNet-S · PSNR + VMAF
Development-stage numbers, being re-run under a stricter reproducibility audit. Salient-region quality is weighted by the model’s own map; no human study yet.
Equal bitratesample.mov · 2000 kbps
+2.49dB
Salient-region PSNR
−0.07
Global VMAF
−0.34dB
Background PSNR
0.24%
Bitrate difference
Side by side
x265 (AQ) · same bitrate
Saliency-guided · same bitrate
Where the model looks
TASED-Net
ViNet-S
Rate sweepsample.mov · TASED-Net
Saliency modelCRF 28
Depth fusionCRF 28 · TASED-Net
Fused · 70% saliency + 30% depth
Depth · near = important