DEV Community

AI-KSK
AI-KSK

Posted on Edited on

MiniMax H3 Long Video in ComfyUI: Per-Shot Review Loops, Digital Humans and Multi-Reference Extension

#comfyui #ai #tutorial #videogeneration

MiniMax H3 Long Video in ComfyUI: Per-Shot Review Loops, Digital Humans and Multi-Reference Extension

Generating longer MiniMax H3 videos is not new; the interesting part of this episode is the review loop. Each shot in a long sequence is reviewed: if it is not good enough, redo it; if it is good, move to the next shot, and finally stitch the accepted shots into one video. The notes below come from the episode subtitles and the published mirror page; no workflow JSON was shipped or parsed, so no node-level parameters are claimed beyond what the episode states.

The three workflows

The core nodes come from a plugin, and the author built three graphs around them, all following a human-in-the-loop logic:

  1. Digital human — image plus audio driven, with unlimited extension.
  2. Image-to-video — with review and extension loops.
  3. Multi-reference — with review and extension loops.

The author states that with 15-second segments a single run can reach 45 seconds or more than a minute.

How the production plan node works

The right side of the graph is a production-plan node, not a plain text box. In it you define scene 1, scene 2, scene 3 and so on — the shot list. Besides per-segment prompts, you can set a global prompt, similar to a prompt relay. The node also holds the render width and height, the duration of each segment (the episode uses 5 seconds), and how many steps each shot runs (8 steps when using an accelerated model).

Prompt conventions follow the official H3 style. Typing @ opens reference options: picture1 and Audio1, which trigger the image and audio inputs. The global prompt describes what the whole video does; each segment prompt describes the action for that shot and can differ per scene. In the episode each scene uses 124 frames as its render length.

Hardware and deployment

The author recommends renting a 5090 or a 4090 48G card so a single segment can be longer and higher resolution. The mirror is deployed on 优云智算:

  1. Open the mirror page: https://www.compshare.cn/images/FP86Uv4w0GMV
  2. Deploy an instance and wait for the image to come up.
  3. Open ComfyUI, pick a workflow, load the image and audio, then run.

Links

Limits

  • The 5-second segment, 8-step and 124-frame values are the episode's settings, not universal defaults.
  • The review loop is manual: quality depends on how you judge each shot.
  • The mirror is a deployment convenience; third-party plugin nodes keep their own licenses.

Top comments (0)