zensim — interactive corpus comparison

← V_X champion site · methodology

Pick corpora, choose any X/Y pair of metrics, get scatter + step-5 line + per-band SROCC. Per the CID22 2023 paper Table 5 band cutoffs. Data is queried in-browser with DuckDB-WASM (HTTP-range over parquet); all CPU work happens on a Web Worker.

Try first: the default pick (AIC-3 corpus, X = q, Y = score_ssim2) shows the canonical sweep-time encoder-quality-vs-perceptual-metric relationship. Switch Y to score_dssim to see how DSSIM ranks pairs differently, or to human_jnd to see how the metric tracks the subjective JND. AIC-4 carries the same shape on a 300-row paper-curated corpus with reconstructed JND.

Corpora

loading manifest…

Axes

Filters

Render

idle

Scatter + step-5 line

Per-band SROCC

BandRange (X) n SROCC KROCC PLCC RMSE
no data yet

Candlestick by X-axis bin (p5–p25–p50–p75–p95)

Each box shows the Y distribution within a 5-unit X bin. Box edges = p25 / p75 (the interquartile range); whiskers = p5 / p95. For AIC-4 the reconstructed-JND CI bounds (`human_jnd_ci_lo`/`hi`) could populate the whiskers directly per-pair — TODO if Y=human_jnd.

Y → codec param lookup

Enter a target Y value; the table lists rows whose Y is within the tolerance, grouped by (codec, version), with their encoded bytes when available. This is the user-facing "I want zensim 70, what should the codec do?" lookup.

codecversion n Y median X median bytes median
no query yet