MiniMax H3 webinar with Luka Tisler: Direct video models

MiniMax H3 is one of the best open video models right now. It makes video and sound together, from a prompt, an image or your own references.

In our 1 October webinar, Luka Tisler from Lighthouse AI Academy showed how he uses it, from first test to finished shot. Here's what he shared. Watch the recording or get his workflow and slides.

“MiniMax H3 is an awesome, amazing model, and if you haven't tried it, you should. Today is all about how to try it and how to work with it.”

Luka Tisler, opening the session

LUKA'S SETTINGS AT A GLANCE

	Test render	Final render
Resolution	0.4 MP (about 480p)	0.8 MP
Steps	4 to 6	35 to 40
Use it to	Check motion and pick a seed	Render the keeper

•Prompts: paste MiniMax's prompt guide into ChatGPT or Claude and let it write them.

•Same character every shot: add a character sheet as a reference.

Luka teaches this in depth at Lighthouse AI Academy: AI for Creative Leaders, a 13-week course for creative directors, studio leads and producers, and Compositor’s Comfy with Doug Hogan of ActionVFX. Course workflows run on Floyo.

What is MiniMax H3?

MiniMax H3 is a video model from MiniMax. Its weights are free to download from Hugging Face, and it works in ComfyUI. Unlike most video models, it creates the sound at the same time as the picture, so dialogue, music and effects match the shot.

	What it does
Clip length	4 to 15 seconds
Sound	Stereo audio in the same pass, 11 languages
Start from	Text, a first and last frame, or references
References	Up to 9 images, 3 videos and 3 voice or audio clips
Also available	H3 Max, a faster closed version you use through an API

Luka's 5 tips for better MiniMax H3 videos

Test small, then go bigger

Preview at about 480p (0.4 megapixels). Render your final at 0.8. Go higher and a clip can take 45 minutes or more, and the model starts ignoring your prompt.

Use few steps for tests, many for finals

4 to 6 steps is enough to see if the motion works. For the final, use 35 to 40 steps.

Kill bad takes early

Add the preview override node. It shows the motion after one step, so you can stop a bad render instead of waiting for it to finish.

Don't reuse prompts from other models

A Seedance prompt won't work well here. Paste MiniMax's prompt guide into your AI assistant and ask it to write H3 prompts for you. The links are in Luka's slides.

Get consistency from references

Make a character sheet of your character from a few angles and add it as a reference. For two characters, put them on one sheet. For a consistent location, add a wide shot of the set.

How to run MiniMax H3: locally or on Floyo

Locally, you need a big GPU. Even the compressed INT8 version of the model is a 32 GB file, before the text encoder and audio files. Luka uses Kijai's compressed versions, which render faster. His slides list every file and where it goes.

On Floyo, there's nothing to install. The workflows run on H100 GPUs with 94 GB of memory, every model is already loaded, and you only pay while a clip is generating.

Try MiniMax H3 on Floyo →

What else you can do with MiniMax H3

•Clone a voice. Luka gave it one photo, a product shot and 5 seconds of his voice. Two minutes later he had a 10-second ad, in his own voice.

“This thing cloned my voice based on 5 seconds of audio. This is insane.”

Luka Tisler, live in the session

•Set the first and last frame. H3 fills in the motion between them. Try it

•Copy motion. Give it a dance video and a character, and your character moves the same way.

•Replace objects. Luka swapped a car for a Shelby Cobra, and H3 even added the shadow.

•Turn clay renders into finished shots. Grey Blender pass in, lit shot out. Try it

•Use LoRAs. There are already LoRAs for speed, camera moves, detail and style.

Good to know: H3 can't output transparent backgrounds (alpha), and clips are 4 seconds minimum.

Luka's advice: keep testing

Asked how to get a consistent scene as the camera moves, Luka shared his reference trick, then the bigger point:

“Testing, testing, testing. R&D, R&D, guys. A lot of R&D. We are scratching the surface on the video models. We don't know what they're capable of yet.”

Luka Tisler, co-founder, Lighthouse AI Academy

His other tip: the best new tools don't trend on social media first. They show up on GitHub and Hugging Face. Luka has an AI agent scan GitHub every week and report which projects are gaining stars, so he finds them before everyone else.

MiniMax H3 FAQ

How do I use MiniMax H3 in ComfyUI?

Update ComfyUI, which has built-in H3 nodes and templates. Download the model files. Or open a ready-made workflow on Floyo with nothing to install: MiniMax H3 image to video.

How much VRAM does MiniMax H3 need?

A lot. The compressed INT8 model alone is 32 GB, and the full model is 62 GB. On Floyo it runs on 94 GB GPUs, so your own hardware doesn't matter.

Does MiniMax H3 make sound?

Yes. It generates stereo audio with the video, in 11 languages, and can copy a voice from a short clip.

How long can MiniMax H3 videos be?

4 to 15 seconds per clip.

What's the best resolution for MiniMax H3?

Test at 0.4 megapixels and finish at 0.8 with 35 to 40 steps.

How do I write MiniMax H3 prompts?

Use MiniMax's own prompt guide. The easiest way is to give the guide to an AI assistant and let it write the prompts.

Can I use MiniMax H3 commercially?

Read the license on Hugging Face first. It doesn't cover the US, EU, UK or South Korea, and very large companies need separate permission.

Try it yourself

“We at Lighthouse Academy are using Floyo for everything Comfy. It's insanely fast. The whole workflow shows up faster than my local workflow.”

Luka Tisler, on why his academy runs on Floyo

NO INSTALLImage and text to videoA start image and a prompt, with video and sound out.NO INSTALLReference to videoYour characters, products and voice.

Watch the full session · Get Luka’s workflow and slides

Want to go deeper? Luka teaches AI for Creative Leaders at Lighthouse AI Academy. See the course

OVERVIEW

MiniMax H3 webinar with Luka Tisler: the ComfyUI settings he uses, how to prompt H3, keep characters consistent and clone a voice. Run it on Floyo, no install.