← mouad el alj / tensor-compression [live] · loading 720p sample…

tensor-compression

A browser-oriented approximate extension of the generalized ℒ-product from my published research (Springer PROMS vol. 497, 2025). Overlapping spatial tiles and exact per-tile SVD (parallel workers) make higher-resolution videos practical in-browser; this is an engineering extension, not a claim of theoretical improvement.

SOURCE Big Buck Bunny 720p sample

open video or LTC2

threshold τ — effective rank —

DECOMPOSITION 128px tiles · 16px overlap · adaptive rank The bundled 720p video uses the same resize path as uploads.

Loading the bundled 720p sample.

source

thresholded reconstruction

PSNR

—

RAW / LTC2

—

GPU RECON

—

CPU RECON

—

reconstruction—

MSE—

relative error—

source filenot a file

raw float32 tensor—

LTC2 at current τ—

source / LTC2—

decomposition—

Metrics compare the reconstruction with the normalized source tensor. File size is the exact uploaded file size; raw size is the extracted float32 tensor.

No upload occurs. Decoding, decomposition, reconstruction, metrics and LTC2 serialization all happen locally in this tab.

how it works

Browser media APIs probe both the bundled 720p sample and uploaded videos, then decode an explicit no-upscale resize choice within 1280×720. Both sources use the same extraction path. The frame budget decreases as resolution rises. An orthonormal DCT couples color and time; overlapping 128px spatial tiles then receive truncated exact SVDs in parallel workers, with a resolution-dependent rank cap. Neighbor-aware feather weights normalize the reconstructed tiles before inverse DCTs.

The threshold applies soft shrinkage max(σ − τ, 0). Components with σ > τ are written to the float16 LTC2 file alongside tiled geometry. The high-quality end of the slider has finer nonlinear control.

The same decomposition also powers the motion-magnification demo, where the temporal-frequency domain the codec stores is used to amplify invisible motion instead of compressing it.

No upload occurs. A reopened LTC2 file can be reconstructed, but it does not contain the original video bytes; its left preview is therefore the decoded factor baseline at the stored threshold, not the original source. Legacy LTC1 files are intentionally rejected.

Video decoding, tiled decomposition, reconstruction, metrics, and LTC2 serialization all happen locally in this tab. ← mouad el alj · email · cv