Introduction

Fifteen seconds is lovely, but stories want minutes. H3MD Director is a fresh open-source ComfyUI node pack from developer YJ that turns MiniMax H3 into a long-form, multi-shot video studio. You plan shots in a visual Director, it generates them segment by segment with a disk cache, and it promises 2+ minutes at 1080p in one run. You can install it today from the Comfy Registry.

Version 0.1.0 was tagged as the first public release on October 31 at 15:12 UTC (23:12 MYT). v0.1.1 reached the Comfy Registry at 18:50 UTC with a security-scan cleanup. The code is AGPL-3.0-or-later with one extra term: keep a visible "Based on H3MD Director" credit. The bundled two-pass engine is GPL-3.0, and the H3 weights aren't included, so they keep their own license.

We covered ComfyUI's core context-windows support for long H3 video on October 30. H3MD takes a different route: shot planning, chained latent continuation and segment repair on top of the core H3 nodes.

What ships

The core chain is three nodes, plus one optional node for motion copy:

  • H3MD Director: a visual editor with an asset bin, timeline, shot tabs and @-mention media tags for each picture, video and audio reference. It writes everything to one plan_json
  • H3MD Loop Sampler + H3MD Loop Options: runs every shot. Shots that are too long for your VRAM get split automatically
  • H3MD Save Video Interp: streams frames to MP4 chunk by chunk, with optional RIFE interpolation up from H3's fixed 24 fps
  • H3MD Depth Control (optional): Depth Anything V2 plus SAM3 person masks drive the Fun ControlNet for motion copy from a reference video

The repo includes two example workflows. The main one, H3MD Director.json, recreates the 133-second motion-copy sample and has model links embedded in it.

How it stays long and stable

Continue in latent space

Each new segment starts from the last 39 frames of the previous segment's final latent, not from decoded pixels. That avoids the slight softening and color shift that pixel round-trips add at every join. Hard cuts (overlapIn 0) start fresh instead.

Two-pass sampling

Each segment runs 6 steps at half resolution. A learned H3 latent upscaler then lifts it to full size, and 3 more steps finish it there. The audio follows one trajectory across both passes.

Re-render only what broke

Spot a glitch in segment 9 of 16? Type 8-10 into rerun_segments and run again. Those segments re-render, and the last block's tail is locked to the cached head of the next segment, so everything else stays untouched. In the author's words: no 5-hour rerun for one glitch.

The drift fix, with numbers

Chained H3 continuations tend to drift: colors deepen, sharpness climbs and faces age. The author traced the color drift to the low-res pass and added lowres_anchor, which pulls each segment's per-channel latent stats back to the first segment of the shot. Measured over a 133-second take, saturation held at 0.33–0.36 with the anchor and rose from 0.34 to 0.48 without it.

Install

You need a ComfyUI build with the core MiniMax H3 nodes (0.43 or newer recommended). No extra pip packages are required:

comfy node install comfyui-h3md-director

You can also search H3MD Director in ComfyUI-Manager. Open the template from Templates → ComfyUI-H3MD-Director, then use Install Missing Custom Nodes to pull in ClipProj, KJNodes, VFI and friends.

Quick start

  • Set stop_after = 1 on the Loop node and render only the first segment to check the look and speed
  • Set it back to 0 with resume_from_cache on. Segment 1 loads from the cache and the rest render
  • Shot lengths follow 5+17n frames at 24 fps, and width and height snap to multiples of 64
  • Write shot prompts with the bundled h3md-prompt skill, either in an agent or pasted into any chat AI

Specs and limits

  • Samples in the README:
    • a 2560×1408, 15-second single segment
    • a 1080p long take of 2:29
    • a multi-shot story of 2:19
    • a motion-copy clip of 2:13
  • VRAM tuning is calibrated on an RTX 5090 32 GB. On cards under 16 GB, the author recommends a lower resolution such as 768×1344, shorter segments and fewer reference media
  • Swapping the 32B text encoder for 4B + ClipProj saves 10+ GB of VRAM, at the cost of slightly weaker prompt understanding
  • The disk cache is about 3 GB per 243-frame 1088×1920 segment (fp16), so make room on a fast drive
  • Reference photos leak into the video: grey studio backdrops turn into grey studio scenes, so use real-location photos
  • It's very early: the repo had 1 star and the registry listed 0 downloads when we checked, so community test reports don't exist yet

Conclusion

H3MD Director is the kind of release I love: a solo builder who measured things, removed a fix that didn't work, and documented why. If you've been stitching H3 clips by hand, this is a well-organized way to plan, chain and repair long takes on your own GPU. Start with one segment, watch the cache fill and let the story run.

Want to feel how H3 moves before going long? Try this on Gen → https://artrealmai.com/gen?utm_source=magazine&utm_campaign=gen&utm_content=h3md-director-long-multi-shot-minimax-h3-comfyui

Original Source

Roll camera, and happy directing!

—Aurelia ♡