CLI

CLI notes (HIP)

./h3 --help is the full list. Flags that change quality, speed, or HIP behaviour:

Flag	Meaning
-d, --model-dir	MiniMax-H3 directory
-p, --prompt	Raw H3 prompt
-o, --output	MP4 (default outputs/h3.mp4)
--width / --height	Output size (defaults 864×480)
--frames / --seconds	Duration; --seconds is at 24 fps
--steps	Euler passes (default 20)
--layers	DiT blocks: 50 exact, 45 fast, 40 aggressive
--reuse	Denoiser reuse: 1 close, 2 fast, 3 aggressive
--token-reduction	Opt-in: pair video tokens in middle DiT blocks (faster long T2VA; visible quality trade; not fox-s2 md5)
--seed	Default 42
--first-frame / --last-frame	FL2VA stills
--ref-image	Ordered Ref2VA image (repeatable)
--ref-image-size	match (default) or max
--ref-video / --ref-silent-video / --ref-audio	Other Ref2VA inputs
--ref-video-audio VIDEO AUDIO	Video + separate soundtrack
--frames-dir	Dump PPM frames
--show	Live denoise preview in supported terminals
--profile	Phase + op-class GPU timing
--ssd-streaming	Stream BF16 DiT layers from SSD (not used in fox benches)

Compile with HIP_ARCH=gfx1151 or HIP_ARCH=gfx90a (not a CLI flag). See Getting started.

--ref-video with an embedded soundtrack needs at least 2 seconds of audio (request --frames 56 or more).

fox-s2 vs fox-fast vs showcase

All three are complete T2VA MP4s (video + audio), not stubs. fox-s2 and fox-fast are 512² / 22 frames (~0.9 s at 24 fps).

• fox-s2: --steps 2 --layers 35 --reuse 1 — HIP A/B and gfx1151 md5.
• fox-fast: --steps 20 --layers 45 --reuse 2 — 11 DiT evals; upstream “first fast video”.
• fox showcase: --steps 20 --layers 50 --reuse 1 — README / wiki gallery fox.

For ~15 s, keep fox-fast knobs and use --seconds 15 at 864×480. Add --token-reduction plus H3_INT8_VAE=1 only when wall clock matters more than the quality path (gfx1151 40 min 46 s → 27 min 3 s, gfx90a 12 min 11 s → 8 min 21 s). See Long video and Performance.