H3MD Director renders 2 to 3 minute MiniMax H3 videos in one ComfyUI run and re-renders only the segments that went wrong
H3MD Director is a free, open-source ComfyUI node pack that plans long, multi-shot MiniMax H3 videos in a visual editor and renders them segment by segment, joined seamlessly, at up to 1080p. Every finished segment is kept on disk, so a bad one can be re-rendered on its own, locked to its neighbours.
What it does
You put photos, a reference video and music into the Director, write a prompt per shot and run. Each continued segment starts from the last 22 or 39 frames of the previous segment's latent rather than decoded pixels, so colour does not shift at the joins, and a low-resolution anchor keeps faces and colour stable over a long single-scene take. Typing segment numbers in rerun_segments re-renders only those, and a motion-copy mode drives the motion from a reference video (depth, SAM 3 and Fun ControlNet) while the look comes from your photos. It ships a prompt skill for coding agents that write the shot prompts.
Results and limits
The README samples include a 2 min 29 s 1080p long take, a 2 min 19 s multi-shot story, a 2 min 13 s motion copy and a 2 min 40 s music video lip-synced to a locked song, plus one 15-second segment at 2560×1408.
Worth knowing
It needs ComfyUI with the core MiniMax H3 nodes (0.43 or newer recommended); no H3 weights are included. Its VRAM estimate is calibrated on an RTX 5090 32 GB, and on cards under 16 GB the author says to use a lower resolution such as 768×1344, shorter segments and fewer reference media. Fast motion can briefly show ghost hands; re-render that segment with a new seed. The code is AGPL-3.0 with a bundled GPL-3.0 sampler. The example workflow's model list includes an abliterated (uncensored) 4B text encoder as the low-VRAM option; the README says the official 32B encoder follows prompts best. MiniMax H3's community licence excludes the EU, UK, South Korea and US.