How Long Can a Clip Get? Duration, Health, and Why Cartoons Don't Help

The question

The previous post benched 2–4 second clips. The obvious follow-ups: how long can a clip get on 12 GB before something breaks — and does content matter? Can a cartoon, with less fine detail to maintain, go longer than a photoreal scene?

Long render times are acceptable; the real question is whether long clips stay healthy. So the sweep below measures cost, and the contact sheets measure health — frames pulled at 0%, 33%, 66% and 100% of each clip, because the first frame always looks fine and decay lives in the later ones.

The sweep: no wall in reach

768×512, 25 steps, both models pushed upward until something broke. Nothing broke:

Model	Frames	Video length	Wall clock	Per second of video	Peak VRAM
LTXV 2B	161	6.7 s	58 s	8.6 s	11.2 GB
LTXV 2B	257	10.7 s	95 s	8.8 s	11.3 GB
LTXV 2B	361	15.0 s	146 s	9.7 s	10.0 GB
Wan 2.2 5B	81	3.4 s	113 s	33.6 s	11.2 GB
Wan 2.2 5B	121	5.0 s	167 s	33.1 s	9.8 GB
Wan 2.2 5B	161	6.7 s	239 s	35.6 s	10.2 GB

Two findings. Cost per second of video is flat — duration scales linearly, no super-linear blowup in the tested range. And VRAM never climbed — the same offload behaviour that made 64K text context a dial rather than a cliff absorbs the growing video latent too. Fifteen seconds of video on a 12 GB card, two and a half minutes of compute, no out-of-memory in sight.

The health evidence

LTXV at 15 seconds — the actual clip, then its frames at 0/33/66/100%:

Identity holds completely: same fox, same forest, no morphing. But watch the clip itself and the problem is obvious — almost nothing moves. What the contact sheet politely called “stability through stillness” is, bluntly, a failure mode, and it deserves a number rather than a euphemism. We added one: mean per-pixel difference between consecutive frames, where a held frame scores ~0.

Clip	Motion score
LTXV 4 s	3.5–4.2
LTXV 6.7 s	4.4
LTXV 10.7 s	1.4
LTXV 15 s	1.6
Wan 2 s	2.4–3.3
Wan 6.7 s	1.5

The pattern is unambiguous: past each model’s comfort zone, motion collapses before identity does. The long clips pass every frame-coherence check precisely because the model quietly stops animating. That is the real single-shot duration wall, and it arrives far earlier than the memory wall — around 7 seconds for LTXV, under 5 for Wan, on this card.

Wan at 6.7 seconds (3.3× its benched length) — clip, then sheet:

Coherent, but doubly degraded: the texture has gone painterly versus the crisp 2-second clip, and its motion score has halved. Long clips don’t fail loudly; they quietly stop being video.

Do cartoons go longer? No — and it cost them

The hypothesis was reasonable: flat-shaded content has no fur to shimmer, so it should drift less. Two halves to the answer.

The hard limit can’t move, and now that’s measured: a cartoon prompt at 361 frames cost 146.1 s against the photoreal clip’s 146.0 s, same VRAM. Memory and compute are set by width × height × frames before the model knows what it’s drawing. Style cannot buy you a longer ceiling.

And the health limit moved the wrong way:

The cartoon fox starts charming, loses its face by two-thirds, and dissolves into smears by the end — while the photoreal clip at the identical length held every frame. Our guess at the mechanism: these models are trained overwhelmingly on real footage, so the temporal prior that keeps a scene coherent is strongest exactly there. A cartoon isn’t easier — it’s off-distribution, and off-distribution is where long generations lose their grip first. (One pair of clips; a hypothesis with the usual n=1 caveat, not a law.)

Verdict

• The memory wall is out of reach, but that is the wrong wall to watch. 15 s renders fine in 2.5 minutes — and scores as functionally static.
• The real single-shot limit is motion collapse, measurable and early: roughly 7 s for LTXV, under 5 s for Wan on this card. Past it, models keep identity by giving up movement — clips that pass every coherence check and fail at being video.
• Getting to genuinely moving 30-second content therefore means composition — many short shots inside the motion-healthy window — which is the next experiment.
• Style doesn’t extend duration — costs are geometry, not content — and stylized content may age worse, not better.
• Contact sheets are the cheap instrument that makes all of this visible; they’re four lines of PIL and they caught everything the single-frame checks missed.

RTX 5070 12 GB, ComfyUI 0.37, 768×512, 25 steps throughout. Sweep and style runs via bench-video.py (--frames, --prompt). Sheets unretouched.