Video summary

MiniMax H3 - как создать непрерывное длинное видео. Обзор наборов нод для ComfyUI + рерайтер промта

Main summary

Key takeaways

Technology

Overview

The video is an overview of Minimax H3 + custom ComfyUI node packs aimed at generating long videos with audio (exceeding the H3 “comfortable” ~15s limit) and improving prompt quality via prompt rewriting.

Main topics / tech concepts

1) What Minimax H3 is (baseline)

  • Multimodal video generation/editing tool that accepts text, images, sound, and video as input.
  • Generates short videos (up to ~15 seconds) in a single pass, with stereo sound, up to about 2K resolution.
  • The creator’s emphasis: this video is about extending H3 beyond short clips in ComfyUI.

Long-video custom node sets (2 workflows)

2) Node pack: MiniMax H3 Long Media

Purpose: Extend ComfyUI workflows so H3 becomes a long-form video generator with sound, not just short clips.

Key implementation ideas / features

  • Uses a multi-clip pipeline: planning → segmentation → continuation generation → latent sequence splicing → final video/audio decode

  • Continuation is built using overlap between chunks, rather than simply appending based on the last frame.

  • Installation: git clone into the ComfyUI custom nodes folder, then restart ComfyUI.

  • Includes an author workflow template in the repo.

Workflow structure (as described)

  • Reference images node: upload 1–4 reference images (for consistency control).
  • Model selection block.
  • Planner node (“Global Prompt” + per-clip prompts):
    • Global prompt affects all clips/scenes.
    • Per-clip prompt defines each segment (scene prompt).
    • Clip duration: set manually in seconds or auto.
    • Supports continuous story or cut-based scenes.
  • Acceleration nodes:
    • Mentions “Sage attention” to speed generation.
  • Long media setup:
    • Reference image wiring and output sizing based on reference.
    • “Workflow Mode” is critical—changing it can drastically alter results (the speaker recounts mixing up modes and getting unexpected output).
  • Sampling block:
    • Sampler/scheduler; default 8 steps mentioned as workable.
    • Presets/tips for different GPU VRAM sizes (8/12/16GB etc.) to avoid OOM.
  • Tiling / tilecoding:
    • Optional, depending on memory needs.

Important tuning / notes

  • Overlap is default ~22 frames.
    • Speaker suggests different overlap tradeoffs could help (e.g., lower for calm/static, higher for heavy motion), though overlap may not be directly user-controllable in this pack.
  • Workflow is actively updated.
  • Quality goal: long output with minimal visible splicing.
  • Recommendation: read the built-in node tips/hints, including RAM/performance presets.

3) Node pack: Continuum (long video via chunk chain)

Purpose: Generate long videos by running a sequence of linked chunks, carrying over intermediate state.

Key implementation ideas / features

  • Runs linked chunks and passes original audio/video latents between them.
  • Optionally supports “CRES Fix” (speed/compatibility enhancement).
  • Mentions a “spectrum” node as an accelerator/speed-up component.
  • Installation: same git clone + restart approach.
  • Includes assembled workflows in the repo (examples folder).

Workflow configuration highlights

  • Mode controls for first/last frames and reference sources:
    • Switches decide whether it uses first/last frame behavior or acts like text-to-video.
  • Reference image/video support:
    • Up to ~3 reference images.
    • Can also use video reference and audio reference for conditioned, potentially lip-synced outputs (speaker explicitly mentions enabling video+audio references).

  • Prompt format via timeline / chunking:

    • Uses a single prompt field with a timeline format.

    • Example concept: “scene 0–5 seconds: … prompt …”

    • Repo provides correct vs incorrect prompt format examples.

    • Chunk timing assumption:
    • At recording time, chunks are equal duration (likely 5s or 10s), though future updates may allow variable chunk lengths.
    • Key parameters:
    • Chunk duration and number of chunks.
    • Overlap frames presets:
      • ~22 normal motion
      • 5 very slow/static
      • 39 strong motion
    • Seed control: random or fixed.
    • Status/output nodes are included, with a reminder to consult the repo for details beyond basics.

Comparison conclusion (speaker’s take)

  • No fundamental “physics” difference: both rely on latent transfer.
  • Differences are mostly about how continuation is parameterized:
    • Long Media: more control per clip (durations and potentially sampling behavior).
    • Continuum: single timeline view, easier to visualize/edit “where everything goes” over time.
  • With matching settings (models/resolution/fps/chunk length/overlap), quality should be approximately similar.

Prompt quality improvement node pack (rewriter)

4) Node pack: Prompt rewriter for Minimax H3

Purpose: Turn a rough video idea into a structured prompt formatted for Minimax H3, optionally conditioned on reference frames/images and duration.

Core workflow idea

  • A central rewriter node generates the final prompt text.
  • Uses reference images and toggle logic to decide generation mode:
    • If reference frames are present, it can rewrite for image-to-video / video-to-video style setups.
    • Speaker references “video text mode” vs frame-driven modes.
    • The system switches which references are enabled and uses first/last frame logic.

UI features

  • Mentions a control panel with clickable icons to enable/disable reference inputs conveniently.

Model loading & selection

  • The rewriter uses a separate LLM model downloaded into ComfyUI’s Models directory, under an lm subfolder.
  • The repo explains which rewriting model to pick depending on model heaviness (speaker follows an MD instruction file in Russian).

Output sizing controls

  • Includes sizing controls such as Maximum new tokens (example: up to 2048) to control prompt length/verbosity.
  • Keep model loaded option:
    • Helpful when iterating/reranking prompts.
    • If doing only one rewrite before generating with Minimax H3, unloading frees GPU RAM for H3.

Value claim

  • Even if it doesn’t perfectly capture the idea, it provides a correct structure and format that Minimax H3 “eats” well (tags/formatting guidance).
  • Speaker recommends using it to learn the correct prompt structure.

Additional product/support mentions (non-core)

  • The creator discusses cloud rental for ComfyUI via “Comfy in the Cloud” (e.g., H100/H200 GPUs, per-second billing, ready images), though the main technical content focuses on the node packs.
  • Channel philosophy:
    • Emphasizes sharing workflows openly
    • Criticizes gated/private prompts in closed groups
    • Encourages starring/open-source support on GitHub

Speakers / sources

  • Main speaker: “Zhenya” (host of the AI Generation channel).
  • Custom node authors / community contributors referenced:
    • Lev Wizard / Valera — author of MiniMax H3 Long Media
    • Wizard — credited within Long Media workflow setup (sampling/performance tips)
    • Alex (“Eatonix” and TG chat moderator mentioned as Alex) — author of the prompt rewriter node pack
  • Sources referenced implicitly: GitHub repositories for the three ComfyUI node packs:
    • MiniMax H3 Long Media
    • Continuum
    • Minimax H3 prompt rewriter (Alex’s pack)

Original video