Added version v2, which improves the audio path.

This is the first complete workflow I have built from scratch.

My goal was to find a practical way to generate MiniMax H3 videos quickly at a low base resolution while still reaching above-HD final output on a 16 GB VRAM GPU. After testing different approaches, this is the most efficient and stable combination I found without turning the process into several separate manual workflows.

With this configuration, I have been able to generate a continuous 40-second Full HD video in approximately 42 minutes on a 16 GB VRAM GPU.

The workflow uses Split Sampling to perform the first denoising steps at a lower resolution, where they are much faster. It then upscales the latent and completes the refinement at higher resolution. Each segment is followed by an MMH3 Ultimate Upscale pass with temporal chunking and tiled diffusion, allowing higher-resolution refinement while keeping VRAM use manageable.

Main Features

• Two sequential MiniMax H3 segments in one workflow.

• Native Motion Context between Segment 1 and Segment 2 for visual and audio continuity.

• Native audio is preserved through the Split Sampler, Motion Context and Ultimate Upscale stages.

• Split Sampler workflow: efficient low-resolution first steps followed by higher-resolution refinement.

• Ultimate Upscale after each segment for cleaner detail and higher final resolution.

• Segment 2 retry mode: regenerate only the second segment from the saved Segment 1 context.

• Multi-reference character workflow through the Deno H3 reference node, allowing an ordered set of character, face, outfit, scene and audio references to be used through one clean input system.

• Visible user controls for prompts, references, seeds, resolution, duration and sampling settings.

• Internal routing, calculations and support nodes are folded to keep the canvas usable.

Important

This is an experimental workflow, but it is prepared to work as provided. The visible controls are intended for normal use; the internal sampler, audio, Motion Context and Ultimate connections should be left unchanged unless you fully understand the pipeline.

The workflow was designed and tested around a 16 GB VRAM setup. Results, speed and stability will still depend on your GPU, installed model versions, LoRAs, resolution and duration.

I'm new to this, so don't be too hard on me if you don't think it's that great lol.