Smooth Workflow Minimax H3

SMOOTH WORKFLOW Minimax H3 Advanced update - 10/23/2026

"Minimax H3 Smooth Workflow Advanced" update: The new version now has options for References. Make sure to fully update your Comfyui and the extensions.

All Images on the showcase used the SmoothMix MinimaxH3 v2.0 lora.

• At the time of making the Showcase: Stability Matrix v2.16.4 and ComfyUI v.0.42.1.

• Added the WhatDreamsCost extension (Check the Extensions list below).

• Added Image, Audio and Video References. Now you have 3 options:

1. Image to Video: use for img2vid, First to Last Frame or Last Frame generations

2. Text to Video: use for txt2vid

3. References to Video: use for all the MinimaxH3 references functions.

• Changed a few of the notes instructions.

• Added Civitai's video generations settings to the "Size Settings Example" note.

SMOOTH WORKFLOW Minimax H3

Minimax H3 gives you a lot of control.

This workflow steps back a little bit to focus on the basics: Img2Vid, First2LastFrame and Txt2Vid.

Remember you can also use just the Last frame if you want.

So, before we begin, all videos on showcase used the following:

• MODELS

diffusion_model

- [minimax_h3_fl2va_pruned_int8_convrot.safetensors] -

vae

- [minimax_h3_video_vae_fp16.safetensors] -

- [minimax_h3_audio_vae_fp32.safetensors] -

text_encoder

- [qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors] -

vae_approx

- [taeh3.safetensors] -

Of course you should use the setup of models/loras/encoders that your hardware can handle and your own preferences. ^^

• EXTENSIONS

After downloading all thats necessary make sure to these extensions installed and updated:

- [WhatDreamsCost] - https://github.com/WhatDreamsCost/WhatDreamsCost-ComfyUI

- [ComfyUI-KJNodes] - https://github.com/kijai/ComfyUI-KJNodes

- [rgthree's ComfyUI Nodes] - https://github.com/rgthree/rgthree-comfy

- [ComfyUI Easy Use] - https://github.com/yolain/ComfyUI-Easy-Use

- [ComfyUI Frame Interpolation] - https://github.com/Fannovel16/ComfyUI-Frame-Interpolation

- [ComfyUI-VideoHelperSuite] - https://github.com/Kosinkadink/ComfyUI-VideoHelperSuite

- [Deno Custom Nodes] - https://github.com/Deno2026/comfyui-deno-custom-nodes

- [MTB Nodes] - https://github.com/melMass/comfy_mtb

- [Adaptative Prompts] -[https://github.com/Alectriciti/comfyui-adaptiveprompts

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

- [Custom Scripts] - https://github.com/pythongosssss/ComfyUI-Custom-Scripts

• Inital Setup

1. Load all you models on the "MODELS" group and make sure to load taeh3 on the "PREVIEW" group. A small graph on the workflow will show you where you should store the models.

2. Choose how you are going to generate the video on "VIDEO SETTINGS" group:

   • 24fps or 48fps: Minimax H3 use 24 fps as base. Using Frame Interpolation will generate a 48 fps video.

   • IMAGE TO VIDEO or TEXT TO VIDEO: Choose the type of generation. If you are using I2V remember to enable the images, First Frame, Last Frame or both on the "IMAGES" group.

   • Video Size: Try to keep the Megapixels (MP) at least on 0.6 to ensure a good quality, regardless of the Aspect Ratio you choose. For High Quality results keep it above 0.8 MP.

     You can check how much MP you are using at the bottom of the Video Size Node.

   • Sampler and Scheduler: The Sampler and Scheduler can affect the image and sound quality of generations - especially if you are using a Turbo Lora. There is no perfect setup. Test what works best for you.

   • Acceleration Nodes: You can enable Sage Attention and Spectrum Minimax H3 to gain a lot of generation speed but it can cost a bit a quality on image and sound.

3. Choose the Loras you are going to use on the "LORAS" group.

4. Type the Prompt for you video on the "PROMPTS" group. It uses Dynamic Prompts. (Check the Dynamic Prompting Guide if you don't know about it). To see the result of you Dynamic prompts or WildCards click on the 'Show Text' node. If you don't know what Dynamic Prompting is don't worry - you can just type the prompt like you normaly would.

5. After you start generating the video you can check the progress of the generation on the "PREVIEW" group. Very useful for long videos - if the video does not look good just cancel it instead of waiting for it to finish,

6. Have fun! ^^