Fifteen Seconds Was the Ceiling — This Node Moves It

MiniMax H3 is a joy for short beats… and a tease for anyone writing a scene. Out of the box it wants roughly fifteen seconds per render. H3-LongVideos, a ComfyUI custom node from Smite79 on Hugging Face, takes that limit as a design problem: turn one written scene into a chain of shots, join them into one continuous video, and keep synchronised audio on the ride.

The card is humming again today (fresh commits, climbing likes). If you already live in H3 + ComfyUI, this is the “make it a short film” button the community kept wiring by hand.

How the Prompt Becomes a Shot List

Workflow is delightfully literary:

  • One paragraph = one shot, blank line between beats
  • The first paragraph is the scene — lighting, place, vibe — prepended to every shot
  • The rest are beats (action + dialogue in double quotes)
  • Optional exact: / exactly: / verbatim: lines push wording to the model word-for-word
  • A character sheet (Name: attributes) keeps people consistent across the chain; pronouns and ages are first-class so references land on the right body

Wire a multiline prompt into the node, set plan_only first (it reports the shot split, lengths, and warnings without burning a render), then go for real. Outputs include finished frames, the soundtrack, and an info dump you should actually read when a shot surprises you.

What It Expects Under the Hood

Drop the folder into ComfyUI/custom_nodes/ and restart the server. Needs ComfyUI 0.31+ with native MiniMax-H3 support (tested on 0.33). Load:

SlotWhat

UNET

a MiniMax-H3 model — hybrid fl2va/ref2va merges work best

CLIP

H3 text encoder (minimax loader type)

VAE

H3 video VAE

audio VAE

H3 audio VAE (the converted one)

Sensible defaults the node now largely owns for you: CFG pinned at 1.0 (H3 is CFG-free), seam trim and silence-on-nonspeech on, and shift_video / shift_audio auto-adjusted when you pair turbo/distill LoRAs with fewer steps so the last step doesn’t invent a half-limb. Pair with a turbo LoRA at ~6–8 steps for speed, or ~20+ without.

Original Source

https://huggingface.co/Smite79/MiniMax-H3-Longvideos

Conclusion

H3-LongVideos is the missing director’s assistant for MiniMax H3 in ComfyUI: scene + beats + character sheet in, chained shots with sync audio out, and an info panel that tells you what the node decided. Pull the latest from Hugging Face, start with plan_only, and give your next H3 idea more than a fifteen-second tease.

Try this on Gen → https://artrealmai.com/gen?utm_source=magazine&utm_campaign=gen&utm_content=h3-longvideos-comfyui-minimax

—Aurelia ♡