What actually breaks a 25-second clip
Not the positions — the schedule
ltx_render_clips samples with DISTILLED_SIGMAS, nine fixed numbers solved
for the clip lengths the distillation used. LTXVScheduler instead derives its
shift from the latent's own token count. A 25-second clip carries about 4.4x
the tokens of a 2-second one, and the fixed schedule does not know that.
The result is a sample that has not converged. On screen it is ghosting — furniture and curtains rendered semi-transparent, overlapping themselves — and roughly half the colour.
| 25 s, 512x320, same prompt and seed | saturation | contrast | ghosting |
|---|---|---|---|
| fixed 8 (what this project shipped) | 0.094 | 0.108 | severe |
| adaptive 8 | 0.127 | 0.124 | gone |
| adaptive 20 | 0.108 | 0.166 | gone, hands and fabric resolved |
| 15 s reference | 0.208 | 0.161 | — |
Final render, 25 s at 1024x640 with the second pass: saturation 0.161, contrast 0.180 — colour close to the fifteen-second reference and contrast above it.
Long clips now switch to the adaptive schedule automatically past 15 s, in
ltx_av_generate. A caller passing steps explicitly still overrides it.
Three times a metric pointed the wrong way
Worth recording, because the same mistake kept recurring in different clothes.
- Laplacian variance said the broken clips had improved 2.5x. They were sharper and broken — fused fingers, blown highlights, collapsing proportions. Sharpness cannot see any of that.
- "Nothing degrades over the clip" came from the same measure: the first-fifth to last-fifth sharpness ratio was flat. Saturation across the whole clip had halved, which no sharpness statistic reports.
- Saturation then preferred adaptive-8 over adaptive-20 (0.127 against 0.108). Side by side, adaptive-20 is clearly better: cleaner hands, defined dress, a legible star belt. Saturation rewards a colour cast as readily as colour.
The common error was choosing a measure before enumerating how the thing could fail. "It will get blurry" was assumed, so sharpness was measured, and the actual failure was colour and anatomy. The user looking at the clips found it in seconds.
Not established
- One prompt, one seed, one duration. Where the fixed schedule starts failing was not bracketed — 15 s is where the switch was placed, from one good clip at 15 s and one bad one at 25 s, not from a sweep.
- 20 steps was picked over 8 by eye. 12, 16, 30 were not tried.
- The two-pass at 25 s does not fit a 16 GB card (OOM at 601 frames); the final render was done on the Space's H200. One pass at 25 s does fit, at 6.58 GiB.
- Position interpolation is still applied above 20 s and still measurably sharpens. It is not what fixed this.