MiniMax H3 Workflows — T2V, I2V, FLF & Ref2V | SageAttention + Turbo LoRA + Upscaler + ControlNet

A collection of MiniMax H3 workflows for ComfyUI, featuring SageAttention, Turbo LoRA acceleration, latent upscaling and refinement.

Included workflows

• FLF: Text-to-Video, Image-to-Video and First–Last–Frame generation.

• Ref2V: Reference-to-Video generation using reference images.

• Ref2V + ControlNet: Reference-image conditioning combined with Canny control from a driving video.

Custom Audio

The FLF and Ref2V workflows support optional custom audio. Load your own audio file and enable Custom Audio to use it during video generation. Disable Custom Audio to let MiniMax H3 generate the audio instead.

The Ref2V + ControlNet workflow can also use the driving video’s audio in the final output.

Main model downloads

Download the H3 diffusion models, standard text encoder, video/audio VAEs, Turbo LoRA and ControlNet weights from:
Comfy-Org / MiniMax-H3

Latent Upscaler

• Upscaler model

• ComfyUI custom nodes

Optional alternative text encoders

The workflows use the standard Comfy-Org text encoder by default. These Heretic variants are optional:

• Heretic Text Encoder — INT8 ConvRot

• Heretic Text Encoder — NVFP4

Setup

Install the required custom nodes, select the appropriate model files, load your references, and enter your prompt. SageAttention is optional. If it is not installed or is incompatible with your setup, you can bypass the SageAttention node.

Some gallery examples use additional LoRAs; their appearance also reflects those models.