H3 MultiRef Update 6: Music Video and AV Workflows

H3 Motion Context MultiRef shipped Update 6 on August 17. The fork of ComfyUI-H3-Motion-Context rebuilds both long-form workflows — Music Video and AV Extension — as checkpoint-free direct-latent chains that stream the final result straight into VideoHelperSuite, removing the large full-movie RAM allocation that older long-form H3 graphs needed.

The bundled music-video demo: six clips generated from two reference images around one master song.

What changed in Update 6

• Checkpoint-free direct-latent chains. Both long-form workflows now pass each sampler's video latent tail straight into the next masked-context node. The Update-5 disk-checkpoint nodes are no longer registered.
• Direct streaming output. New final-output nodes decode one clip at a time and stream completed frames into VideoHelperSuite instead of materializing one complete final IMAGE tensor — the known high-RAM path is no longer exposed.
• Exact AV timing. Final audio stitching uses absolute timeline sample boundaries, avoiding one-sample seam drift at rates such as 44.1 kHz. Small H3 audio-grid undershoots are time-conformed instead of padded with a silence tail.
• Controller-owned bypass. Real ComfyUI bypass mode for the generation and preview groups, so neither current workflow requires rgthree for group control.
• Forward compatibility. Recognizes both the pre- and post-ComfyUI-PR-#15375 H3 mask integration layouts, and the local fallback retires itself once the current layout is present.

Music Video

The 20-slot Music Video workflow builds a multi-clip video around one song. The starter clip is generated with MiniMaxH3ReferenceToVideo, then each clip passes its video latent tail directly to the next. The master song stays authoritative: every clip receives the exact absolute master-song slice in its H3 audio latent with audio mask 0, and the final output uses the original master waveform untouched.

AV Extension

The AV Extension workflow continues a video across multiple H3 generations. One controller drives a shared live chain for existing-video, T2V, and I2V/Custom-Keyframes starts, with two image-reference slots and two optional audio references feeding the starter and all extensions. Protected context snaps to the shared H3 video/audio boundaries (39 / 90 / 141 / 192…), audio keeps a hard preserved region plus a configurable half-cosine feather (8 audio-latent ticks, 0.2 s, by default), and the result streams into VideoHelperSuite without allocating a full final movie tensor.

Example references

The music-video demo ships with two reference images and a song in example_workflows/assets/.

Reference image 1	Reference image 2

Availability

Install by cloning into ComfyUI/custom_nodes/ and restarting:

cd ComfyUI/custom_nodes
git clone https://github.com/seitanism/ComfyUI-H3-Motion-Context-MultiRef.git

Some included workflows expect VideoHelperSuite, KJNodes and rgthree; install any missing node packs and restart. The example music-video assets (two images and one song) go into your ComfyUI input/ folder before running. Note that Update-5 checkpoint nodes are no longer registered — workflows saved against them must be migrated to the current direct-latent versions.