RunningHub MiniMax H3 Dual Sigma Sampler (Legacy):
This node performs the sampling stage for MiniMax H3 audio-and-video generation. It progressively refines an audio/video latent representation using separate noise schedules for video and audio, which a standard KSampler cannot express for H3. It supports task-aware conditioning, video-and-audio generation or audio generation from an existing video, and optional acceleration. The result is still a latent representation rather than a finished video or audio file, so you can pass it to the appropriate downstream decoding workflow.
RunningHub MiniMax H3 Dual Sigma Sampler (Legacy) Input Parameters:
h3_model
The required MiniMax H3 direct model (MINIMAX_H3_DIRECT_MODEL) supplies the transformer used during sampling. Connect a compatible H3 model; for newer task-aware conditioning, the node checks that the model matches the requested task and latent target. No numeric minimum, maximum, or default is specified.
conditioning
The required H3 conditioning (MINIMAX_H3_CONDITIONING) provides the generation guidance. Task-aware conditioning also identifies the task, allowing the node to check compatibility with the model and latent before sampling. No default or numeric range is specified.
av_latent
The required H3 audio/video latent (MINIMAX_H3_AV_LATENT) provides the target latent structure to sample. Its target must be compatible with the conditioning and model; when generating audio from video, its video portion must contain the intended clean video rather than an empty, all-zero video. No default or numeric range is specified.
seed
An integer used to seed sampling, so you can repeat a run with the same settings or change the seed to explore another result. The default is 42; the minimum is 0 and the maximum is 9223372036854775807.
sigma_points
The number of points in the sampling noise schedule. More points generally mean more sampling work: the source notes that 50 sigma points produce 49 DiT forward passes in the stated Euler setup. The minimum is 2 and the maximum is 1000. The default is the source constant H3_DEFAULT_SIGMA_POINTS; its numeric value is not shown in the provided source.
video_shift
A floating-point setting passed to the sampler to shape the video noise schedule. It lets you adjust the video side separately from audio, but the provided source does not define a general visual effect for increasing or decreasing it. The minimum is 0.01, the maximum is 100.0, and the adjustment step is 0.01. The default is H3_DEFAULT_VIDEO_SHIFT, whose numeric value is not shown.
audio_shift
A floating-point setting passed to the sampler to shape the audio noise schedule independently of video. The provided source does not establish a general sound-quality effect for increasing or decreasing it, so compare results against a known-good setup when tuning it. The minimum is 0.01, the maximum is 100.0, and the adjustment step is 0.01. The default is H3_DEFAULT_AUDIO_SHIFT, whose numeric value is not shown.
accel
Selects an optional single-GPU approximation for faster sampling. The tooltip describes off as disabling acceleration; auto as preferring a velocity-cache profile and otherwise trying Cache-DiT; minimax-h3-velocity-cache-v1 as a recommended single-GPU profile using fewer DiT evaluations; minimax-h3-cache-v1 as a Cache-DiT profile that requires cache-dit; and manual-* modes as manually tuned alternatives. The default is off (ACCEL_OFF). The complete set of selectable strings is not available because the referenced choices constant is not included. The source says the acceleration profiles were validated only at 1344×768, 124 frames, 50 steps, and shifts 12 and 3; do not treat accelerated results as ground truth.
denoise_video
Controls whether sampling denoises video as well as audio. The default is True. Set it to False for video-to-audio (V2A): the existing video is treated as a clean visual condition at the video timestep floor, while only audio is denoised. This mode requires a T2VA-layout latent whose video is not empty or all zeros; the source suggests populating it through Encode Video → AV Latent or Separate AV Latent.
cache_dit_rdt
An optional residual threshold used only when accel is manual-cache-dit. A larger value is described as faster but rougher. The minimum is 0.0, the maximum is 1.0, and the adjustment step is 0.01. The default is CACHE_DIT_RDT_COOKBOOK; its numeric value is not shown.
cache_dit_mc
An optional limit on consecutive cached steps, used only with manual-cache-dit. It bounds how long that mode can continue using cached results before refreshing them. The minimum is 1 and the maximum is 32. The default is CACHE_DIT_MC; its numeric value is not shown.
cache_dit_warmup
An optional number of warmup steps for manual-cache-dit. It controls that mode’s initial warmup period rather than the overall number of sampling points. The minimum is 0 and the maximum is 64. The default is CACHE_DIT_WARMUP; its numeric value is not shown.
velocity_stride
An optional DiT refresh interval used only with manual-velocity acceleration. A value of 1 means exact sampling without velocity caching; larger intervals refresh less often. The minimum is 1 and the maximum is 32. The default is VELOCITY_STRIDE; its numeric value is not shown.
sampler_mode
Selects the sampling method. euler is described as the official first-order method, with sigma_points=50 noted in its tooltip. res_multistep is a second-order multistep exponential integrator; the tooltip recommends sigma_points=21 for 20 DiT passes and describes quality roughly comparable to Euler at 50 points with an approximately 2.5× speedup in its stated comparison. The default is euler (SAMPLER_MODE_EULER). The complete choices list is not visible in the provided source. By default, res_multistep forces acceleration off because the supplied acceleration profiles were calibrated for Euler at 50 points.
allow_accel_with_res_multistep
An optional boolean that permits acceleration to run alongside res_multistep. The default is False. The source warns that this combination is not calibrated and noticeably degrades results at sigma_points=21; it says higher point counts would also be needed and that, at comparable quality, it is not more economical than Euler with velocity acceleration.
sampler_config
An optional linked MINIMAX_H3_SAMPLER_CONFIG containing sampler configuration. Leave it disconnected to use the sampler’s defaults. The node passes a connected configuration to the sampling implementation, but the provided source does not show its fields or how they override the other settings; consult that configuration node before changing its controls. No numeric range or explicit value default is specified.
RunningHub MiniMax H3 Dual Sigma Sampler (Legacy) Output Parameters:
sampled_av_latent
The sampled H3 audio/video latent (MINIMAX_H3_AV_LATENT). This is the refined latent result to pass to a compatible downstream decoder or other H3 latent-processing node—not a directly playable media file. Depending on the task and settings, sampling may refine both modalities or generate audio while using an existing video as clean conditioning. The node also records sampling telemetry and model residency information in the output data; for task-aware conditioning, it adds schema, partition, and compatibility fingerprints and validates the resulting latent.
RunningHub MiniMax H3 Dual Sigma Sampler (Legacy) Usage Tips:
- For a baseline comparison, start with
sampler_mode=eulerandaccel=off. Keep the seed fixed while changing one setting at a time so differences are easier to judge. - To try fewer DiT evaluations, use
sampler_mode=res_multistepwithsigma_points=21and leaveallow_accel_with_res_multistep=False. The node’s tooltip recommends this combination but presents its quality and speed comparison as guidance, not a guarantee for every project. - On a single GPU, consider the
minimax-h3-velocity-cache-v1acceleration profile when it is available. Test it against an unaccelerated result for your own resolution, frame count, and shifts, since the stated profile validation covers only one configuration. - For video-to-audio work, set
denoise_video=Falseand supply a T2VA-layout latent containing real video. An empty video latent is not a substitute for visual conditioning. - Use
cache_dit_rdt,cache_dit_mc, andcache_dit_warmuponly when tuningmanual-cache-dit; usevelocity_strideonly formanual-velocity. Start with the provided defaults and change one control at a time.
RunningHub MiniMax H3 Dual Sigma Sampler (Legacy) Common Errors and Solutions:
Missing required input: h3_model, conditioning, or av_latent
- Explanation: Sampling needs all three connected inputs. This is a likely workflow validation message, not a literal error string established by the supplied class.
- Solution: Connect a compatible H3 direct model, H3 conditioning, and H3 audio/video latent to their matching sockets.
Input value is outside the allowed range
- Explanation: The interface restricts values such as
sigma_pointsto2–1000, both shifts to0.01–100.0, andseedto0–9223372036854775807. The exact validation message depends on the workflow host. - Solution: Return the affected setting to its permitted range, or reset it to its displayed default.
Conditioning, model, and latent are incompatible
- Explanation: For task-aware conditioning, the node checks that the selected model and latent target match the conditioning’s task. The exact exception text comes from validation functions not included in the provided source.
- Solution: Use a model and latent created for the same H3 task as the conditioning. Recreate the latent from the intended target if you have mixed components from different workflows.
Invalid H3 audio/video latent
- Explanation: The node validates the supplied latent before sampling and validates task-aware output afterward. A malformed or mismatched latent can therefore stop the run; the validator’s literal error message is not available here.
- Solution: Recreate the latent with a compatible H3 node and check that its target and task match the conditioning rather than editing or mixing latent data manually.
V2A requires a non-empty video latent
- Explanation: With
denoise_video=False, the source requires a T2VA-layout latent containing real video, not an empty all-zero video. The exact runtime message is not shown. - Solution: Populate the video portion using Encode Video → AV Latent or Separate AV Latent, or turn
denoise_videoback on if you intend to generate video too.
Cache-DiT dependency or acceleration setup failure
- Explanation: The
minimax-h3-cache-v1profile is documented as requiringcache-dit. Its installation and failure messages are outside the supplied node source. - Solution: Install and configure the required dependency for your environment, or switch
acceltooffor an available velocity-cache mode.
Processing interrupted
- Explanation: The node checks for cancellation before and during sampling, so stopping a run can raise an interruption message. The precise wording is supplied by the runtime.
- Solution: Restart the run if the interruption was accidental; otherwise, no parameter change is needed.
Out of memory while loading or sampling
- Explanation: The node loads the transformer for sampling and monitors memory use, but the exact memory error depends on the runtime and device. Large generation targets may exceed available GPU memory.
- Solution: Reduce the target size or frame count in the upstream workflow, free GPU memory used by other applications, or use a model-loading setup suited to your available hardware.
