MiniMax H3 Open Weights - Reference to Video

ABOUT THE WORKFLOW

Edit a Clip With a Reference Give the model a reference still, a clip to work from, and a line saying what to change. MiniMax H3 rebuilds the clip with your edit applied, holding the original motion and scene, and writes the picture and the stereo sound in the same pass.

Model

• MiniMax H3 by MiniMax, the lab behind the Hailuo video line, also called Hailuo 3.0. This workflow runs the reference to video checkpoint, which locks a character, object, style, motion, or voice from your references into the clip it builds.

HOW IT WORKS

Step 1. Load your reference picture A still of what you want placed into the clip. Front facing and well lit gives the closest match. Works great with: faces · characters · products · costumes

Step 2. Load the clip you are editing Its motion, scene, and timing become the base the model rebuilds from.

Step 3. Write the change Describe only what changes, not the whole scene. "Replace the woman in the library with the old man in the reference picture, keep the camera move."

Step 4. Set your aspect ratio and clip length Pick a shape on the resolution selector and type a length in seconds. Length comes from that setting, not from the clip you loaded.

Step 5. Hit run and download You get one video file per run at 24 frames per second with stereo sound, saved under video/MiniMax_H3. Ready for: Premiere · DaVinci Resolve · CapCut · After Effects

First time? Leave every setting as-is. The defaults (16:9 · 5 seconds · reference size match · random seed) are the right starting point for almost everyone.

RECOMMENDED SETTINGS

Quick-start guide. Find the goal that matches yours and copy the settings.

• Standard edit (most people) — 16:9 · 5 seconds · reference size match · random seed. The right starting point for almost everyone.

• Faster runs at the same quality — Drop megapixels to about 1.0. The open weights are built around a 768 pixel short edge, so the 2 default sits well above their range and costs time without adding detail.

• Stronger likeness on a face or a product — Switch reference size from match to max. That holds your reference at up to a 2048 pixel short edge for a closer match, and it slows the run because those reference tokens stay active through every sampling step.

• Point the prompt at a specific reference — Tag your inputs in the order you connected them, so your still is Picture 1 and your clip is Video 1, then say what each one is for. Naming the job of each reference lands the edit more reliably than describing it loosely.

• Bring the source clip's sound in — The video loader has an audio output sitting free. Connect it to the reference video audio slot and the model can work from the original voices and room tone. Sound is not pulled off the clip on its own.

• Your output is shorter than your source clip — Length comes from the duration value, not from the video you loaded. Raise duration to cover more of the source, up to the 15 seconds the model handles.

• Repeat a take you liked — Set a fixed seed number instead of leaving it on random. The same references, prompt, and seed give you the same clip back.

• The edit is not landing — Name only the thing that changes and leave the rest of the scene out of the prompt. A cluttered instruction pulls the model away from the source clip.

• Clip length looks slightly off — Frame counts snap to a fixed ladder, so a 5 second request runs 124 frames, which is about 5.2 seconds at 24 fps. That is expected.

Prompt: Say what changes and what stays. "Replace the woman in the library with the man in the reference picture, keep the shelves, the lighting, and the camera move" gives you more than "put the old man in the video." Give every reference one job, and leave the parts of the scene you want untouched out of the sentence.

LEARN

📹 Videos

• Intro to Floyo

• ComfyUI 101 Free Course ft. Sebastian Kamph

• Floyo 101 for Team Collaboration

✨ Quick links

• Help docs

• LoRA Training guide

• Discover more workflows

USE CASES

🎭 Character Replacement Drop a different actor, model, or character into footage you already shot while the blocking and camera move hold.

🛍️ Product Placement Put a specific product into an existing clip so one piece of footage covers several SKUs or several markets.

🎬 Reshoots Without a Camera Change who or what is in a scene when a reshoot is out of budget or the location is gone.

🎨 Series Consistency Keep the same face, costume, or product look across a run of clips by feeding the same reference into each one.

WHAT WORKS BEST / WHAT TO AVOID

✅ Works great

• Front facing, well-lit reference stills of one subject

• Source clips with steady, readable motion

• Prompts that name one change and nothing else

• Giving each reference a single job

⚠️ May produce softer results

• Megapixel values pushed well past the 768 pixel short edge

• Reference stills at an extreme angle or in heavy shadow

• Source clips with fast cuts or several subjects moving at once

• Prompts that redescribe the whole scene alongside the edit

FAQ

What is reference to video in MiniMax H3? Reference to video takes a text prompt plus reference images, video clips, and audio, then generates a new clip that carries the identity, style, motion, camera move, or voice from those references. It is the mode used for swapping a character or object into existing footage, as opposed to animating a still from scratch.

How many references can MiniMax H3 take? The model accepts up to 9 reference images, 3 reference videos, each of which can carry its own soundtrack, and 3 standalone audio clips. This workflow opens with one image slot and one video slot connected, and more slots appear as you wire references in.

What is the difference between reference to video and image to video? They run different weights. Image to video uses the FL2VA checkpoint, which animates from a first frame and optionally a last frame. Reference to video uses the Ref2VA checkpoint, which conditions on a set of references instead of a starting frame. Use image to video to bring a still to life, and reference to video to change what is in a clip you already have.

Does the output keep the original video's audio? Not on its own. The model writes fresh stereo sound in the same pass as the picture. To have it work from the source, connect the audio output on the video loader to the reference video audio slot, because audio is not pulled off a reference clip automatically.

Is MiniMax H3 open source? The weights went public on Hugging Face on 3 August 2026, but the license is not a standard open-source license. The MiniMax H3 Community License covers commercial use free for organizations under 20 million USD in yearly revenue, requires products built on it to display "MiniMax H3" in the interface, and excludes the US, EU, UK, and South Korea from its applicable territory for local deployment. Read the license text before you build on it.

What does it take to run MiniMax H3 reference to video locally? Reference to video needs its own diffusion checkpoint on top of the text and image path, around 21 GB more, which brings the full local package to roughly 63 GB. Community reports put 16GB of VRAM at workable with quantized files and offloading, and 24GB to 32GB at comfortable, with 64GB or more of system RAM once weights start moving off the card.

How to run MiniMax H3 Reference to Video online? You can run MiniMax H3 Reference to Video online through Floyo. No installation, no setup, no 63GB download. Open the workflow in your browser, load a reference and a clip, describe the change, and hit run. Free to try.

WHY FLOYO?

Floyo is the only platform with team collaboration for ComfyUI in the browser. You run workflows with no install. You share run history, assets, and models across your team. You pay only when you generate. Floyo supports open-source and closed-source models.

A designer runs an edit and likes the result. A teammate opens that exact run from shared history and keeps going. No file handoffs. No version confusion.

For studios and enterprise teams, Floyo adds private workspaces, pooled resources, and a team usage dashboard. Other ComfyUI cloud tools run for one person at a time. Floyo runs for the whole team, with transparent per-generation costs.

Ready to try it? Load a reference picture and the clip you want to change, write the edit, and run it. The settings are already set.

→ Launch Workflow, Free

Questions? Watch the free course or check the FAQ above.