SMD provides separate four-step and eight-step adapters for MiniMax H3's reference-conditioned video-and-audio partition. References define the appearance of subjects; the model generates a new scene and soundtrack around them. The practical choice is a matching adapter, schedule and runtime, rather than lowering the steps on an arbitrary H3 graph.
This guide reviews the authors' repositories and a Hugging Face demonstration on October 24, 2026. We have not run the models. The cover is an AI-generated illustration, and published speed or memory figures below belong to their stated sources.
Choose one complete execution route
The four-step model card and eight-step card each provide a student adapter, standalone inference code and a reproducible example. The original Python runner uses Linux, Python 3.12 and two CUDA GPUs with FSDP2; its documented test used two A100 80 GB cards. It is not the single-GPU installation shortcut.
For native ComfyUI, choose the corresponding converted repository: four steps or eight steps. For a browser trial, the SMD Space provides a separate hosted demonstration. These routes share the intended student identity but use different loading and hardware arrangements.
Install the native ComfyUI package
Follow the selected converted repository's setup instructions in your ComfyUI Python environment. Its download script obtains and verifies the required models, copies example inputs and installs the full workflow in the user library. No SMD custom-node checkout is required for that documented native graph.
Open Workflows, expand SMD and select the matching H3-Ref2VA-4step-Compact or eight-step workflow. Alternatively, drag its workflow.json onto the canvas. Confirm that the graph includes both the LoRA loader and video saving before running the supplied example.
Do not substitute H3 weights into the ordinary image-generation template. An error about audio_scale, for example, can indicate that the required audio-video graph is incomplete. Resolve the graph and component identities before diagnosing the prompt. Switch students by installing the corresponding repository's adapter and workflow together.
Budget storage and memory separately
The converted packages offer two profiles. Compact lists approximately 42.0 GB of downloads; the unpruned BF16 transformer profile lists 101.2 GB. Shared components are reused. The compact set combines a pruned INT8 transformer, NVFP4-AWQ text encoder and INT8 video VAE; the BF16-transformer profile still uses a quantized text encoder.
The documented compact test ran 1344×768, 124 frames at 24 fps on an A100 80 GB with a reduced memory budget. Its observed peak was about 24.3 GiB. The authors explicitly say physical 24 GB consumer cards were not tested. Neither a reduced software budget nor the word “Compact” establishes a universal consumer-GPU fit.
The standalone runner's files have a different storage arrangement. Avoid combining a standalone download estimate with the native ComfyUI profile's memory observation as though they described one installation.
Preserve the student schedule
The converted workflow specifies LoRA strength 1, Euler, BasicGuider without CFG and supplied ManualSigmas. Its schedule uses video/audio shifts of 12 and 3, with raw final-interval scale 0.5. Keep those settings with the matching four- or eight-step student; do not add Lightning or Turbo adapters to the baseline.
The original release's sampling contract is part of the model recipe. A step widget alone cannot reproduce its time grid. Keep a saved unmodified graph, then change one creative input at a time after the example completes.
Use references as appearance inputs
Replace the example images in their Load Image nodes and preserve their order in the prompt. A simple exercise is a character reference plus an object reference, with a new location specified in text. Describe one interaction and its sound rather than expecting the reference images to dictate exact motion.
For example, adapt this untested editorial brief to the graph's numbered picture syntax:
The adult explorer from Picture 1 sets the small brass telescope from Picture 2 on a wooden table. A steady medium shot shows the explorer unfold its legs and look toward the horizon. Soft wind, a quiet metallic click and no dialogue.
Inspect both identity and interaction. Then add more complex speech or movement only after the basic binding is correct. The published 124-frame example is about 5.17 seconds at 24 fps, so distinguish the nominal duration label from the delivered frame count.
Understand the browser demo's boundaries
The Space's application source exposes one to three reference images, selectable students and two-to-ten-second duration choices snapped to a frame grid. It uses a pruned base and quantized conditioner on ZeroGPU hardware. Its availability and cold placement time are separate from sampling speed.
Community preference for SMD over PTD in a few clips is useful qualitative evidence, not a controlled ranking. Compare your own complete scenes, including speech and audio, before drawing conclusions. H3Vid's Ref2VA generator is a separate online workflow; neither it nor the PTD article implies that SMD is installed behind the form.