javawock7618's picture
Update README.md
611159c verified
|
Raw
History Blame Contribute Delete
10.9 kB
metadata
tags:
  - comfyui
  - minimax-h3
  - first-last-frame
  - flf
  - reference-to-video
  - r2v
  - video-generation
  - int8
  - sol-attn
  - spectrum
  - turbo-lora
  - lightx2v
  - motion-context
  - latent-upscaler
  - workflow
  - optimization
  - quantization

MiniMax-H3 INT8 FLF & Reference Video Workflows Collection

A collection of optimized ComfyUI workflows for MiniMax-H3, designed for high-speed and controlled video generation using INT8 quantization, Sol-Attn, Spectrum acceleration, Lightx2v LoRA, MiniMax-H3-Turbo-Lora (Experimental), Motion Context, and Latent Upscaling.

These workflows provide optimized pipelines for:

  • First-Last-Frame (FLF) video generation
  • Reference-to-Video (R2V) generation
  • Motion Context-enhanced FLF generation
  • High-speed INT8 inference
  • LoRA-based acceleration
  • Latent video upscaling

Optimized for the ComfyUI-compatible MiniMax-H3 model package.

Base model:

Comfy-Org/MiniMax-H3
https://huggingface.co/Comfy-Org/MiniMax-H3


Workflows

Workflow File Description
First-Last-Frame (FLF) MiniMax_int8_I2V-javanoYYMM.x.json Generate controlled video transitions using first and last frame images, with optional Motion Context conditioning.
Reference-to-Video (R2V) MiniMax_int8_R2V-javanoYYMM.x.json Generate videos while maintaining reference image characteristics.

Features

INT8 Optimization

Uses the INT8 version of MiniMax-H3.

Benefits:

  • Reduced VRAM usage
  • Faster inference
  • Improved memory efficiency
  • Optimized local ComfyUI generation

Sol-Attn Acceleration

Integrated support for:

ComfyUI-SolAttn_triton
https://github.com/kijai/ComfyUI-SolAttn_triton

Sol-Attn provides Triton-based optimized attention kernels for ComfyUI.

Benefits:

  • Faster attention computation
  • Reduced attention overhead
  • Improved generation throughput

Spectrum Acceleration

Integrated support for:

ComfyUI-Spectrum-MiniMax-H3
https://github.com/xmarre/ComfyUI-Spectrum-MiniMax-H3

Features:

  • Faster sampling
  • Reduced generation time
  • Optimized MiniMax-H3 inference
  • Native ComfyUI workflow integration

Motion Context

Available in the FLF workflow only.

ComfyUI-H3-Motion-Context
https://github.com/NikoDemon80/ComfyUI-H3-Motion-Context

Motion Context provides additional temporal motion information to the MiniMax-H3 FLF generation pipeline.

Features:

  • Motion-context conditioning
  • Additional temporal motion guidance
  • Motion-aware FLF generation
  • Improved control over generated movement

This custom node is currently integrated into:

MiniMax_int8_I2V-javanoYYMM.x.json

It is not used by the R2V workflow.


Lightx2v LoRA

Integrated support for the Lightx2v LoRA for MiniMax-H3:

Kijai/MiniMax-H3_comfy
https://huggingface.co/Kijai/MiniMax-H3_comfy

Lightx2v LoRA is used as an additional acceleration method for reducing the number of sampling steps required for video generation.

Features:

  • Reduced sampling steps
  • Faster video generation
  • Accelerated workflow iteration
  • Compatible with MiniMax-H3 ComfyUI workflows

The optimal LoRA strength and sampling configuration may vary depending on the workflow and desired output quality.


MiniMax-H3-Turbo-Lora (Experimental)

Optional support for:

MiniMax-H3-Turbo-Lora-ComfyUI
https://huggingface.co/drbaph/MiniMax-H3-Turbo-Lora-ComfyUI

Turbo-LoRA enables faster generation using fewer sampling steps.

Features:

  • Low-step generation
  • Faster previews
  • Rapid workflow testing
  • ComfyUI-compatible implementation

Recommended for:

  • Fast prototyping
  • Motion testing
  • Preview rendering

This feature is experimental. Output quality may vary depending on:

  • Sampling steps
  • LoRA strength
  • Prompt complexity
  • Motion intensity
  • Resolution settings

Latent Upscaler

Integrated support for:

ComfyUI-MiniMaxH3_LatentUpscaler
https://github.com/Tr1dae/ComfyUI-MiniMaxH3_LatentUpscaler

The Latent Upscaler provides a dedicated post-generation upscaling pipeline for MiniMax-H3 video output.

Features:

  • Latent-space video upscaling
  • Higher output resolution
  • Improved fine details
  • Enhanced texture and image quality
  • Designed specifically for MiniMax-H3 workflows

The upscaling stage can be used after the initial video generation to improve output resolution and visual detail without requiring the entire generation process to run at the higher resolution.


Workflow Details

First-Last-Frame (FLF)

MiniMax_int8_I2V-javanoYYMM.x.json

A First-Last-Frame video generation workflow using MiniMax-H3 INT8.

Input:

  • First frame (starting image)
  • Last frame (ending image)
  • Optional Motion Context

Generates smooth transitions between two keyframes.

Features:

  • First Frame + Last Frame conditioning
  • Controlled motion transitions
  • Scene transformation
  • Pose changes
  • Camera movement control
  • Optional Motion Context conditioning
  • INT8 inference
  • Sol-Attn acceleration
  • Spectrum acceleration
  • Lightx2v LoRA support
  • Optional Turbo-LoRA support
  • Latent Upscaler support

Recommended for:

  • Cinematic transitions
  • Character animation
  • Camera movement
  • Transformation effects
  • Keyframe-based storytelling
  • Motion-guided generation

Reference-to-Video (R2V)

MiniMax_int8_R2V-javanoYYMM.x.json

Reference-guided video generation workflow.

Features:

  • Reference image conditioning
  • Character consistency
  • Appearance preservation
  • INT8 inference
  • Sol-Attn acceleration
  • Spectrum acceleration
  • Lightx2v LoRA support
  • Optional Turbo-LoRA support
  • Latent Upscaler support

Recommended for:

  • Character videos
  • Product visualization
  • Consistent subjects
  • Image-driven animation

Motion Context is not included in this workflow.


Required Custom Nodes

Install the following custom nodes:

ComfyUI-SolAttn_triton

https://github.com/kijai/ComfyUI-SolAttn_triton

ComfyUI-Spectrum-MiniMax-H3

https://github.com/xmarre/ComfyUI-Spectrum-MiniMax-H3

ComfyUI-H3-Motion-Context

https://github.com/NikoDemon80/ComfyUI-H3-Motion-Context

ComfyUI-MiniMaxH3_LatentUpscaler

https://github.com/Tr1dae/ComfyUI-MiniMaxH3_LatentUpscaler

Install via:

ComfyUI Manager → Install Missing Custom Nodes


Required Models

Main Model

ComfyUI-compatible MiniMax-H3:

https://huggingface.co/Comfy-Org/MiniMax-H3

Lightx2v LoRA

Kijai/MiniMax-H3_comfy

https://huggingface.co/Kijai/MiniMax-H3_comfy

Used for Lightx2v acceleration.

Optional Turbo-LoRA

MiniMax-H3-Turbo-Lora-ComfyUI

https://huggingface.co/drbaph/MiniMax-H3-Turbo-Lora-ComfyUI

Used for experimental low-step acceleration.


Recommended Environment

  • Latest ComfyUI
  • Latest ComfyUI-Manager
  • Comfy-Org/MiniMax-H3
  • ComfyUI-SolAttn_triton
  • ComfyUI-Spectrum-MiniMax-H3
  • ComfyUI-H3-Motion-Context
  • ComfyUI-MiniMaxH3_LatentUpscaler
  • Kijai/MiniMax-H3_comfy
  • MiniMax-H3-Turbo-Lora-ComfyUI (Optional)

Installation

  1. Update ComfyUI and ComfyUI-Manager.

  2. Install the required custom nodes:

ComfyUI-SolAttn_triton
https://github.com/kijai/ComfyUI-SolAttn_triton

ComfyUI-Spectrum-MiniMax-H3
https://github.com/xmarre/ComfyUI-Spectrum-MiniMax-H3

ComfyUI-H3-Motion-Context
https://github.com/NikoDemon80/ComfyUI-H3-Motion-Context

ComfyUI-MiniMaxH3_LatentUpscaler
https://github.com/Tr1dae/ComfyUI-MiniMaxH3_LatentUpscaler

  1. Download the main model:

Comfy-Org/MiniMax-H3
https://huggingface.co/Comfy-Org/MiniMax-H3

  1. Download the Lightx2v LoRA:

Kijai/MiniMax-H3_comfy
https://huggingface.co/Kijai/MiniMax-H3_comfy

  1. Optional: download the experimental Turbo-LoRA:

MiniMax-H3-Turbo-Lora-ComfyUI
https://huggingface.co/drbaph/MiniMax-H3-Turbo-Lora-ComfyUI

  1. Load the workflow:

FLF:

MiniMax_int8_I2V-javanoYYMM.x.json

R2V:

MiniMax_int8_R2V-javanoYYMM.x.json


Performance Configuration

The workflows support multiple acceleration technologies that can be combined depending on the desired balance between speed, quality, motion control, and stability.

Standard Fast Configuration

INT8 + Sol-Attn + Spectrum

Accelerated Configuration

INT8 + Sol-Attn + Spectrum + Lightx2v LoRA

FLF Motion Configuration

INT8 + Sol-Attn + Spectrum + Motion Context

Experimental Maximum-Speed Configuration

INT8 + Sol-Attn + Spectrum + Lightx2v LoRA + Turbo-LoRA

Latent Upscaler can then be used as a post-generation stage when higher output resolution is required.

Performance depends on:

  • GPU architecture
  • VRAM capacity
  • Resolution
  • Frame count
  • Sampling steps
  • LoRA strength
  • Motion Context settings
  • CUDA/driver version

Recommended Workflow

For a balance between speed and quality:

1. Generate at the base resolution
Use the INT8 MiniMax-H3 workflow with Sol-Attn and Spectrum.

2. Apply Motion Context when using FLF
Use Motion Context when additional temporal motion guidance is required. This feature is available in the FLF workflow only.

3. Apply acceleration LoRA
Use Lightx2v LoRA when fewer sampling steps are desired.

4. Optional Turbo-LoRA
Use Turbo-LoRA for experimental ultra-low-step generation and rapid previews.

5. Latent Upscaling
Use ComfyUI-MiniMaxH3_LatentUpscaler after generation when higher resolution or additional detail is required.


Notes

  • FLF = First-Last-Frame video generation.
  • R2V = Reference-to-Video generation.
  • Motion Context is currently available only in the FLF workflow.
  • Lightx2v LoRA is used for accelerated generation.
  • Turbo-LoRA is experimental.
  • Sol-Attn and Spectrum acceleration depend on hardware and software compatibility.
  • Latent upscaling increases output resolution after the initial generation stage.
  • Output quality may vary depending on sampling steps, LoRA strength, resolution, Motion Context settings, and acceleration configuration.
  • Combining multiple acceleration methods may require careful tuning of sampling steps and LoRA weights.

Additional Resources

MiniMax-H3

https://huggingface.co/Comfy-Org/MiniMax-H3

Sol-Attn

https://github.com/kijai/ComfyUI-SolAttn_triton

Spectrum MiniMax-H3

https://github.com/xmarre/ComfyUI-Spectrum-MiniMax-H3

H3 Motion Context

https://github.com/NikoDemon80/ComfyUI-H3-Motion-Context

Lightx2v LoRA

https://huggingface.co/Kijai/MiniMax-H3_comfy

MiniMax-H3-Turbo-Lora-ComfyUI

https://huggingface.co/drbaph/MiniMax-H3-Turbo-Lora-ComfyUI

MiniMax-H3 Latent Upscaler

https://github.com/Tr1dae/ComfyUI-MiniMaxH3_LatentUpscaler


License

Please follow the licenses of:

  • Comfy-Org/MiniMax-H3
  • ComfyUI
  • ComfyUI-SolAttn_triton
  • ComfyUI-Spectrum-MiniMax-H3
  • ComfyUI-H3-Motion-Context
  • ComfyUI-MiniMaxH3_LatentUpscaler
  • Kijai/MiniMax-H3_comfy
  • MiniMax-H3-Turbo-Lora-ComfyUI
  • Any additional custom nodes, LoRA, or third-party models used with these workflows.