Long video

Long T2VA (10 s and 15 s)

MiniMax-H3 on h3-hip.c supports up to 362 aligned frames (~15 s at 24 fps). Long clips are dominated by DiT denoise wall time, not the short fox presets. Same knobs on both ISAs: --steps 20 --layers 45 --reuse 2 (11 DiT evaluations), 864×480, T2VA, no references.

Build with HIP_ARCH=gfx1151, HIP_ARCH=gfx90a, or HIP_ARCH=gfx942 (Getting started).

Showcase clips

Clip	GPU	Duration	E2E wall	Denoise wall	File
10 s cinematic office	gfx1151	10.1 s (243 f)	24.7 min	21.2 min	long-10s-cinematic.mp4
15 s cinematic office	gfx1151	15.1 s (362 f)	40 min 46 s	36 min 38 s	long-15s-cinematic.mp4
15 s all-opts	gfx1151	15.1 s (362 f)	27 min 3 s	23 min 10 s	sampler+TR+INT8 VAE; long-15s-all-opts-2026-09-03.log
15 s cinematic office	gfx90a	15.1 s (362 f)	12 min 11 s	10 min 47 s	main 2026-09-02; 480 px VAE tiles
15 s + --token-reduction	gfx90a	15.1 s (362 f)	8 min 21 s	6 min 50 s	opt-in; TOKEN_REDUCTION.md

Posters: assets/showcase/long-*-cinematic.jpg
Halo phase splits: docs/perf-runs/LONG_VIDEO.md
MI210 session: docs/perf-mi210/STATUS.md

10 s was not re-timed on MI210 or on main gfx1151. The gallery MP4 is the quality path (no TR). gfx1151 quality path on main (2026-09-03): E2E 40 min 46 s, denoise 36 min 38 s, VAE 174 s (2×1 @ 480 px), DiT peak 27.9 GiB. Log: long-15s-default-2026-09-03.log.

Reproduce the 15 s clip

MODEL=/path/to/MiniMax-H3

./h3 --profile -d "$MODEL" \
  -p "15 seconds, 16:9 landscape cinematic. A lone software engineer works late in a dim home office lit only by monitor glow and a desk lamp. Photoreal live-action feel with subtle handheld camera breathing.

[0–3 seconds] Medium shot from behind the desk. Code scrolls on dual monitors; warm red accent light reflects on glass. Ambient: quiet keyboard clicks, soft fan hum, distant city rain.

[3–6 seconds] Slow push-in over the shoulder. On screen, glowing matrix tiles and magenta wavefronts visualize a neural network training. The engineer pauses, sips coffee. Sound: gentle electronic pulse, a single soft notification chime.

[6–9 seconds] Cut to close-up of hands typing, then rack focus to a small window showing a red fox walking through digital snow inside the monitor reflection. Sound: rising synthesized tone, subtle wind.

[9–12 seconds] Smooth lateral move across the desk: terminal windows, GPU metrics, and a grid of video frames assembling on screen. Warm amber grade, volumetric dust in the lamp beam.

[12–15 seconds] Controlled pullback reveals the full workspace at rest. The engineer leans back, satisfied. Sound: clean final impact, room tone fades.

No readable text, no logos, no subtitles. Premium technology documentary aesthetic." \
  --width 864 --height 480 --seconds 15 \
  --steps 20 --layers 45 --reuse 2 --seed 42 \
  -o outputs/long-15s-cinematic.mp4

Add --token-reduction and H3_INT8_VAE=1 when wall clock matters more than the quality path. GPU sampler is not required on gfx1151.

	quality path	--token-reduction
gfx1151 15 s E2E	40 min 46 s (main 2026-09-03)	27 min 3 s (all-opts)
gfx90a 15 s E2E	12 min 11 s	8 min 21 s (denoise 6 min 56 s all-opts)

Reproduce the 10 s clip

Same prompt structure with [0–3] … [8–10] blocks and --seconds 10. See docs/perf-runs/LONG_VIDEO.md for the full command and --profile phase table.

Frame alignment

Output frame count must align to 5 + 17×N (5 … 362). --seconds N picks the nearest valid count at 24 fps (--seconds 15 → 362 frames, 15.08 s).

Project page

Live clips and the reproduce block: alexhegit.github.io/h3-hip.c/#long