ComfyUI MiniMax H3 Timeline Director

ComfyUI MiniMax H3 Timeline Director

Start with the MiniMaxH3 All-in-One Full Timeline Director workflow: one lightweight graph covers text-to-video, image reference, audio-driven generation, video editing, character replacement, motion transfer, digital humans, and manually segmented long-video generation.

简体中文 · English

Headline feature: lightweight unlimited-length video generation

The plugin splits any target duration into continuous segments and completes them in one ComfyUI execution: per-segment generation with one shared seed, direct continuation from the previous AV-latent tail, adaptive Drift-Control video masking, Soft AV audio continuity, overlap removal, and final synchronized assembly. Extend the result by increasing the segment count—without generic Loop nodes or duplicated sampler chains. Practical length is limited only by local VRAM, RAM, disk space, and ComfyUI execution limits.

For long-form digital humans, standalone audio can be set to Locked Original Audio. The source waveform is sliced on the timeline and injected through the native AV-latent mask/sigma path, so lip motion remains audio-driven while the final soundtrack preserves the uploaded recording unchanged.

Download the all-in-one workflow · Chinese guide · Chinese prompt specification

Lightweight unlimited-length workflow

SelfLift two-stage sampling: fast 75% low-res + 25% high-res generation

The Material Planner now integrates a MiniMax H3 adaptation of SelfLift progressive-resolution sampling. Enable Two-stage sampling, select an H3 Latent Upscaler from ComfyUI/models/latent_upscale_models/, and set High-resolution sampling steps. A recommended starting point is 25% of the scheduler's total steps: for an 8-step schedule, set 2 high-resolution steps to run 6 low-resolution steps + latent lift + 2 high-resolution steps. The high-resolution value must be greater than zero and lower than the scheduler's total step count.

Two-stage toggle, model selector, and high-resolution step control

This is not a per-segment upscale shortcut. Every later segment carries both the preceding native low-resolution latent tail and the final high-resolution context, with Drift-Control applied at both resolution stages. The resulting long-video path avoids progressive quality loss from repeated resizing or VAE round trips and removes the blur, white flashes, and visible seams normally associated with multi-segment generation.

Creator benchmark: a 1536×832, 29-second video rendered directly in approximately 10 minutes with a 75% low-resolution / 25% high-resolution schedule. Actual speed depends on GPU, VRAM, model, step count, and reference complexity. When standalone audio is locked or reference-video source audio is enabled, the native zero-denoise AV path restores one continuous source waveform; creator tests retain 99%+ content and timing consistency. The final MP4 saver may still re-encode audio.

One workflow covers text-to-video, image-to-video, audio-reference generation, image plus audio, reference-video editing, character replacement, motion transfer, digital-human lip sync, and multi-segment long videos without visible seams or progressive degradation.

Two directly generated, approximately one-minute examples

Both videos were produced in one plugin execution and are 52.625 seconds / 1263 frames / 24fps. Click a poster to play or download the original MP4. All media is hosted as GitHub Release assets, so it adds nothing to the plugin clone or installation size.

| Finite direct-latent continuation | Reference