MATLOWAI/H3-Dials

H3 Dials: training-free steering for MiniMax H3

Dials are ~14 KB steering files made from prompt pairs, with no training. Each dial is one direction in MiniMax H3's processed-prompt space. Its strength is a slider:

• negative moves every generation toward the pair's A side;
• positive moves it toward the B side.

Make and apply them with the ComfyUI nodes: github.com/matlowai/ComfyUI-Dials.

On Pika H3 Flex, with de-rope

Top to bottom: MiniMax H3 base, 25 steps, no de-rope; then Pika H3 Flex (6 steps) with motion +0.5 and de-rope, at dread 0, +0.5 and +1. Same prompt and seed. A de-roped Pika clip takes about as long as one 25-step base render (about 6 min at 1280x704 on one RTX PRO 6000 Max-Q) and comes out with more motion and crisper frames; the end card has the timings, motion and sharpness.

Motion: Freeze ↔ Zoomies

Left to right: dial −2, base, dial +2 (same prompt and seed).

file	mode	load with	notes
dials/h3_motion.safetensors	hook (after H3's text refiner)	ComfyUI-Dials Load Dial → Apply Dial	the reference version. −2: near-frozen, cinemagraph-like (0.05× base motion). +2: everything zooms (≈2.7× base motion)
dials/h3_motion_lora.safetensors	LoRA file (condition_proj bias diff)	stock Load LoRA (model only), or Load Dial	the same pairs, about 0.6× the effect per unit (use ~1.6× strength)

Built from 24 prompt pairs: 8 subjects × 3 phrasings of still/posing/frozen vs running/jumping/moving energetically. Pair agreement is 0.78. The pairs are in example_pairs/.

Villainy pack: eight flavours of evil (hook dials)

Made with debias on, so each flavour is its own direction (motion is unrelated to all of them). +1 to +2 is the sweet spot; +4 is extreme.

file	turns the scene toward	notes
dials/h3_evil_sinister.safetensors	menace: shadows across faces, dim rooms, looming silhouettes	⚠️ tends to break the audio: speech gets crushed (Whisper word error 0.84 at +2 vs 0.00 at base). Keep it low for clear dialogue
dials/h3_evil_creepy.safetensors	uncanny: unnervingly wide smiles, subtly wrong movement, sterile spaces
dials/h3_evil_cold.safetensors	icy: flat, calculating, emotionless
dials/h3_evil_unhinged.safetensors	maniacal, wild-eyed	⚠️ really is unhinged, beware: screaming, lunging intimidation even at +2 (leans aggressive; no gore seen)
dials/h3_evil_seductive.safetensors	a charming smile hiding dark intent	speech stays clean
dials/h3_evil_corrupted.safetensors	decay and rot creeping in
dials/h3_evil_dread.safetensors	foreboding atmosphere, no character needed
dials/h3_evil_impish.safetensors	mischievous trickster	the lighter end

These have been checked on a handful of scenes, so treat them as creative tools. Pair files are in example_pairs/.

On Pika H3 Flex (6 steps)

Dread 0 / +0.5 / +1 with motion +0.5 on Pika H3 Flex (same prompt and seed; audio from +1). The dials work on Pika unchanged and keep the prompt and characters. On Pika, +0.5 to +1 is the useful range for motion and evil alike. Ready-made workflows (Pika + motion dial + evil slot, with and without de-rope): ComfyUI-Dials › Pika H3 Flex + Dials.

Motion dial on Pika: 0 to +3

1280x704, same prompt and seed. +0.5 / +1.0: same scene, more action (+20 % / +23 % camera-removed motion). +1.5 / +2.0: the scene restages and characters morph. +3: loses the prompt.

De-rope vs no de-rope

Motion +1.0. De-rope finds stretches that move too fast for the frame rate, re-renders them slowed down and maps them back to real time, keeping the original audio. The workflow default re-enters at inject 0.48; a 6+6 schedule is included as an experimental workflow. At 1280x704, a de-roped Pika clip takes about 5.5–6 min, about the same as one 25-step render of the base model without de-rope (about 5.5 min).

These two videos were rendered with an earlier wording of the example prompt, before the magical-girl rewrite. Re-renders with the current wording are coming.

What makes a dial different

• Not a LoRA or slider LoRA: nothing is trained. A dial is computed from prompt pairs in about a minute and only steers what H3 already knows.
• Not a RefMod: no reference images or video. A dial changes how much of a quality any prompt gets, not what is shown.
• Not an SD-era embedding: it isn't a word you type. It's a signed direction added to every prompt token, where the DiT reads the prompt.
• Not a negative prompt: H3 is guidance-distilled (no CFG), so negative prompts don't work. Dials give that kind of control back, in both directions.

There's a full comparison table in the GitHub README.

Thanks

Thanks to DaveSnow, whose Natural_detailV2 dataset and every-second-block observation started the research behind this. Powered by MiniMax H3.

License

These dial weight files are Model Derivatives of MiniMax H3 and are distributed under the MiniMax H3 Community License Agreement (LICENSE-MiniMax-H3-Community.txt, notice in NOTICE), including its Applicable Territory and Acceptable Use Policy terms. The ComfyUI-Dials node code is separate and AGPL-3.0, on GitHub.