javawock7618/comfy-MiniMax-H3-workflows

MiniMax-H3 INT8 FLF, R2V, Bridge & TTS Workflows Collection

Optimized ComfyUI workflows for MiniMax-H3, featuring INT8 inference, native Block Sparse Attention, Veda Sparse Attention, Spectrum, Lightx2v Turbo LoRA, native Extend, Fun ControlNet Union V2V control, Face Refine, TTS, Latent Upscaler, Bridge, and Group Bypass.

Veda Sparse Attention is supported by both the I2V and R2V workflows as an optional attention optimization.

These workflows are designed to provide flexible local video and audio generation, with optional acceleration and refinement stages that can be enabled or bypassed as needed.

Workflows

Workflow	File	Description
First-Last-Frame (FLF)	MiniMax_int8_I2V-javanoYYMM.x.json	FLF/I2V video generation with Extend, ControlNet Union V2V, Latent Upscaler, and optional Veda Sparse Attention.
Reference-to-Video (R2V)	MiniMax_int8_R2V-javanoYYMM.x.json	Reference-guided video generation with Reference LoRA, ControlNet Union V2V, Extend, Latent Upscaler, and optional Veda Sparse Attention.
Bridge	MiniMax_int8_Bridge-javanoYYMM.x.json	Generates a transition between Video A and Video B.
Face Refine	MiniMax_int8_FR-javanoYYMM.x.json	Face refinement for generated video.
Text-to-Speech (TTS)	MiniMax_int8_TTS-javanoYYMM.x.json	Audio-only speech generation.

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

Features

INT8 Inference

The workflows use INT8 MiniMax-H3 model weights to reduce memory requirements and improve local inference efficiency.

• Reduced model memory requirements compared with higher-precision weights.
• Improved memory efficiency for local inference.
• Suitable for systems with limited VRAM relative to full-precision configurations.

Actual generation speed and output quality depend on the GPU, attention backend, sampling configuration, and enabled optimizations.

Native Block Sparse Attention

The FLF / I2V and R2V workflows use ComfyUI's native Block Sparse Attention node for optimized attention processing.

This is separate from Veda Sparse Attention, which is an additional optional attention optimization.

Native Block Sparse Attention does not require an additional custom node.

Spectrum

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

Spectrum provides additional MiniMax-H3 sampling optimization.

The workflows support Spectrum alongside other compatible acceleration methods. Sampling steps and LoRA strengths may require adjustment to balance speed and quality.

Lightx2v Turbo LoRA

Model: https://huggingface.co/lightx2v/Minimax-h3-Turbo

An acceleration LoRA intended to reduce the number of sampling steps.

The appropriate LoRA strength and sampling configuration depend on the selected workflow and output requirements.

Veda Sparse Attention — I2V and R2V

Model: https://huggingface.co/Veda-Sparse/Minimax-H3-R2VA-Veda-Preview

ComfyUI implementation: https://github.com/veda-sparse/Veda-on-ComfyUI

Veda Sparse Attention is an optional attention optimization for MiniMax-H3 video generation. It is available in both the I2V and R2V workflows.

It uses a learned predictor to identify relevant visual attention tiles and reduce attention computation.

• Available for both I2V and R2V workflows.
• Requires the Veda-on-ComfyUI custom node and a compatible predictor checkpoint.
• Separate from the native ComfyUI Block Sparse Attention implementation.
• Does not replace the MiniMax-H3 base model or the R2V Reference LoRA.
• Optional; the workflows can be used without enabling Veda Sparse Attention.

The published R2VA Preview is an experimental configuration associated with a Turbo-LoRA setup. Compatibility with other model, LoRA, and sampling configurations should not be assumed.

Follow the official repository instructions for installation, model placement, and supported inference backends.

Native Extend

MiniMax-H3 supports native video extension for continuing existing video content.

The workflows use native Extend functionality rather than requiring the former Motion Context custom node.

Fun ControlNet Union — V2V Control

Model repository: https://huggingface.co/Kijai/MiniMax-H3-experimental/tree/main/controlnet

Integrated into the I2V and R2V workflows.

Supported control types:

• OpenPose
• Canny
• Depth

Available models:

• minimax_h3_fun_controlnet_union_pruned_bf16.safetensors
• minimax_h3_fun_controlnet_union_pruned_int8_convrot.safetensors

Use the model variant compatible with your workflow and inference configuration.

Latent Upscaler

Custom node: https://github.com/LBH-123-AI/Comfyui_Minimax_h3_latent_Upscaler

Model: https://huggingface.co/LBH-123-AI/Minimax_h3_latent_Upscaler

Integrated into I2V and R2V.

Generation pipeline:

Low-resolution generation → Latent Upscaling → High-resolution decoding

The workflow uses a low-resolution generation stage followed by latent-space upscaling and refinement.

Bridge

Generates a transition between two existing clips:

Video A → Generated Bridge → Video B

Designed to create intermediate footage connecting two video segments.

Face Refine

Custom node: https://github.com/Carasibana/ComfyUI-H3-FaceRefine

Provides automatic face detection and refinement for generated video.

Text-to-Speech

MiniMax-H3 can generate audio without running the video generation path.

The TTS workflow is intended for audio-only speech generation.

Group Bypass

Custom node: https://github.com/Oratorian/ComfyUI-Switchboard

Used to enable or bypass optional workflow groups, including optional processing and optimization stages.

Workflow Details

First-Last-Frame (FLF)

MiniMax_int8_I2V-javanoYYMM.x.json

• First + Last Frame conditioning
• Fun ControlNet Union / V2V
• OpenPose / Canny / Depth control
• Native Extend
• Latent Upscaler
• Native Block Sparse Attention
• INT8 inference
• Spectrum
• Lightx2v Turbo LoRA
• Optional Veda Sparse Attention
• Group Bypass

Veda Sparse Attention is optional and can be enabled when the compatible Veda implementation and predictor are installed.

Reference-to-Video (R2V)

MiniMax_int8_R2V-javanoYYMM.x.json

• Reference image conditioning
• Reference LoRA
• Fun ControlNet Union / V2V
• OpenPose / Canny / Depth control
• Native Extend
• Latent Upscaler
• Native Block Sparse Attention
• INT8 inference
• Spectrum
• Lightx2v Turbo LoRA
• Optional Veda Sparse Attention
• Group Bypass

Required Reference LoRA:

minimax_h3_ref_lora_rank_256_bf16.safetensors

Download from: https://huggingface.co/Kijai/MiniMax-H3-experimental/tree/main/loras

Veda Sparse Attention is an optional acceleration component for the R2V workflow.

Bridge

MiniMax_int8_Bridge-javanoYYMM.x.json

Video A → Generated Bridge → Video B

Creates a generated transition between two existing video clips.

Face Refine

MiniMax_int8_FR-javanoYYMM.x.json

Uses ComfyUI-H3-FaceRefine for automatic face refinement.

Text-to-Speech

MiniMax_int8_TTS-javanoYYMM.x.json

Audio-only MiniMax-H3 speech generation.

Workflow Compatibility

Feature	FLF / I2V	R2V
INT8 inference	Yes	Yes
Native Block Sparse Attention	Yes	Yes
Spectrum	Yes	Yes
Lightx2v Turbo LoRA	Yes	Yes
Native Extend	Yes	Yes
Fun ControlNet Union	Yes	Yes
Latent Upscaler	Yes	Yes
Veda Sparse Attention	Optional	Optional
Reference LoRA	No	Required

Required Custom Nodes

Install the custom nodes required by the workflows you intend to use.

Custom node	Purpose
ComfyUI-Spectrum-MiniMax-H3	Spectrum sampling optimization
ComfyUI-H3-FaceRefine	Face refinement
Comfyui_Minimax_h3_latent_Upscaler	Latent upscaling
ComfyUI-Switchboard	Group Bypass
Veda-on-ComfyUI	Optional Veda Sparse Attention for I2V and R2V

Native Block Sparse Attention is provided by ComfyUI and requires no additional custom node.

Veda-on-ComfyUI is required only when enabling Veda Sparse Attention. It is not required for standard I2V / FLF, R2V, Bridge, Face Refine, or TTS generation.

Required Models

Main Model

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

Use the model variant compatible with the workflow's INT8 loader.

R2V Reference LoRA

`minimax_h3_ref_lora_rank_256_bf16.s