De-rope A/B — LTX-2.5 against H3, on the shared player

flipbook · this page · the H3 arm comparison

Pick a scene, then click any row to load it against the reference. F flip · drag to wipe · Space play · ←→ frame · Esc close. One pair is live at a time; the previous one is torn down.

staircase (hold map) is the same ruler under an arbitrary clock, read from the hold-map sidecar the arm saved (H3 Save Hold Map), so oracle-shaped arms on any model get it; staircase is read on the dilated pass and is the LTX de-rope verdict: 1.00 = the model handed the frame-repeat init straight back, 0.00 = it spread the motion across the hold. Renormalised by dilation, so d=8 and d=16 compare. alt (motion alternation) is the only column an H3 row and an LTX row share: roping holds then jumps, so lower is smoother. It is within-scene only — tempest's baseline sits at +0.370 and swordspin's at +0.960 because they are differently roped clips, and on panrun it is not a verdict at all, because under a lateral pan the camera owns the motion-energy curve. detail and psnr are against each pass's own reference, within scene and resolution; a starred psnr is against the baseline resized to the arm's canvas (hover for the note). Every timing here is on a QUANTISED base — int8 + convrot on both sides (H3 fl2va_pruned_int8_convrot 21.0 GB, LTX 22b-distilled…int8-convrot) — so none of it is comparable to numbers quoted elsewhere against a full-precision model. Peaks are also with --gpu-only, which keeps the nvfp4 text encoder resident; without it the same arms sit far lower.
Every LTX arm is a WINDOW, not the whole clip. The DiT's time axis is 20 s, so at d=16 the The LTX arms on this deck run a FLAT clock: the H3 jerk oracle chooses where the 57-frame window sits (its hot span), and inside it every frame is held exactly d times. The H3 arms are oracle-clocked (per-frame holds of 1-4 with ramps and bridges). An oracle-shaped LTX clock, the H3 hold map scaled onto LTX's 8-frame latent blocks, is in progress and is not what these rows measure. The largest legal source window is 57 frames (913 dilated = 19.02 s); 73 frames would be 24.35 s and the minter refuses it. Swordspin arms are 57 frames from 16 (of 73), panrun arms 57 frames from 39 (of 124, the parasol burst). The default reference is the source cut to that same window; clicking a full-length arm (124 / 73 frames) switches the reference to the full-length source or baseline, so an A/B starts on the same frame - picking the full-length baseline instead will look misaligned, and that is the windowing, not a broken render.
Scene Pass Show Reference Compare