RunningHub MiniMax H3 Dual Sigma Sampler (Legacy)

This is the heart of the whole MiniMax H3 pack - the node that actually runs the DiT and denoises your clip. Everything else is setup; this is where the 33B omni-modal model does its thing, generating video and its native stereo audio together. And it can't be a standard KSampler, because H3 runs two latent streams (a 24-channel video stream and a 32-channel audio stream) with different noise schedules. Hence "dual sigma": separate shift values for each stream.

The inputs that matter

• sigma_points - 50 by default, and per the tooltip that means 49 actual DiT forwards. This is your "steps." Fewer points = faster, rougher; the res_multistep mode below is the clever way to get fewer forwards without the quality hit.
• video_shift / audio_shift - 12 and 3 by default. These set how the noise schedule skews for each stream. They're the knobs you're allowed to touch but probably shouldn't until you're deep in the weeds; the defaults are what the pack and the reference implementation were tuned around.
• denoise_video - the one that changes what you're doing. True (default) denoises video and audio together: a normal T2VA/FL2VA/Ref2VA run. Set it to False and you're in V2A mode: the video latent is treated as a clean visual condition (timestep floor) and only the audio gets denoised. That's how you add sound to existing footage - feed the frames in via RHMiniMaxH3EncodeVideoAVLatent or SeparateAVLatent first.
• accel - single-card speedups. off is off; auto prefers the velocity-cache profile; minimax-h3-velocity-cache-v1 cuts DiT forwards (the recommended single-GPU path, no extra packages); minimax-h3-cache-v1 is Cache-DiT and needs pip install "cache-dit>=1.3.0". The tooltip's warning is honest: these profiles were only validated at 1344×768 / 124f / 50 steps / shift 12·3, so treat them as "works there," not as ground truth everywhere.

The speed lever that's actually worth it

sampler_mode = res_multistep is the hidden gem. The default euler follows the official first-order recipe (50 points). res_multistep is the second-order multi-step exponential integrator - the same one in ComfyUI's own H3 template - and the pack's claim is sigma_points=21 (20 DiT forwards) lands near euler-50 quality at roughly 2.5× the speed. That's the single best quality/effort trade you can make in this pack. Note it currently forces accel=off, and allow_accel_with_res_multistep exists but is flagged as uncalibrated - leave it off.

Output

sampled_av_latent - the denoised AV latent, which is not viewable yet. It goes to RHMiniMaxH3DecodeAV (with the VAE bundle) to become actual frames and an audio waveform. A common newbie mistake is stopping at the sampler and wondering where the video is.

Install

cd ComfyUI/custom_nodes
git clone https://github.com/RH-RunningHub/ComfyUI-RH-MiniMax-H3.git
pip install -r ComfyUI-RH-MiniMax-H3/requirements.txt

Restart, pull the ~95 GiB bundle into ComfyUI/models/MiniMax-H3-INT8-CONVROT/, and remember this sampler is where VRAM and patience go to be tested. The license reminder still applies - H3's Community License excludes the US, EU, UK and South Korea from local-weights use.