Simply Advanced - MiniMax H3

Simply Advanced - MiniMax H3 Workflow by altoiddealer

This workflow was built from the ground up by myself working daily over the course of 2-3 weeks, aiming to be well structured, expandable, and easy to use without sacrificing any functionality.

New in v1.5:

Seamless Continuation has been replaced with Seamless Extension

1. Slices off X frames from your video input(s) to set Video/Audio guide(s).

2. MiniMax H3 generates Video/Audio which begins and/or ends with those sliced inputs.

3. The workflow stitches the result with the original video input(s) using configurable overlap blending strategies. Optional: enter the subgraph to enable saving the raw generation result.

Seamless Extension modes:

• Seamless Continuation (generate from input video as start point)

• Seamless Convergence (generate into input video as end point)

• Can be used in combination (Generate between two input videos)

Main features:

• Supports up to: 6 Images, 2 Videos, 3 Audio.

  • ANY can be set as reference, guide, or both. This is much more useful than the average user understands - I have not yet seen this support in any other workflow.

• Latent Upscale and/or 2-Stage Generation

  • Intuitive settings/LoRA management for each stage.

• 1-4 steps LLM Prompt Enhancement

  • Subgraph contains a dashboard to generate / edit LLM's responses before they pass into the next stage.

  • Includes my personal 4-Step Ref2VA pipeline.

    • Step 1 Prompt Enhancement/Expansion: (Optional Step). Allows the LLM to interpret the user's prompt more verbosely and in relation to the reference materials. Catching/editing discrepancies here will yield much better results in next steps.

    • Step 2: Reference Analysis: Creates a detailed mapping of the references in relation to the Expanded Prompt. The resulting Reference Map is included as part of the context in the next steps.

    • Step 3: Creative Director: Considers Expanded Prompt + Reference Analysis to create a Creative Blueprint.

    • Step 4: Prompt Compiler: Conforms the Reference Map + Creative Blueprint into the H3 prompt structure

• Includes my personal 1-Step FL2VA Prompt Enhance

  • The system prompt begins with base instructions, adding task specific instructions (T2VA / I2VA / FL2VA / L2VA)

  • The mode is detected automatically based on the state of Image 1 / Image 2.

• You, the user, can easily add your own LLM instructions into this system as selectable options. Take a shot at making your own multi-stage pipeline!

• Seamless video extension option - continuation idea taken from Seed Hunter workflow - later expanded to include convergence/combination usage.

• Force Audio option - idea taken from Seed Hunter workflow

• Easily expandable framework

  • Easily add/manage speedups, LoRAs, etc

• Is the poster child for an amazing new node, Pass or None. Workflow developers, start using this!!

ComfyUI Requirements:

• Be updated.

• MUST have comfyui-frontend version < v1.50.5 (Fixed broken Custom Combo nodes in subgraphs)

Required Custom Nodes:

• ComfyUI-Easy-Use

• ComfyUI-KJNodes

• comfyui_essential-er (My repo). Was optional, now Required for v1.5+ for its superior video/audio merging node

Optional Custom Nodes:

• Load Video (Upload) from ComfyUI-VideoHelperSuite Feel free to replace with another Load Video node.

• MiniMax H3 Latent Upscaler

• MiniMax H3 RefMods

• Datetime String from comfyui-various - Useful for naming outputs. NOT REQUIRED

• ComfyUI-Spectrum-MiniMax-H3 - Speedup option. NOT REQUIRED