MiniMax H3 Max Editing Nodes in ComfyUI: Recast, Extend, Reframe, Insert, Lip Sync and Camera Control

ComfyUI now has six new MiniMax H3 Max nodes that edit video you already have, rather than making a new clip from scratch.

You can swap the cast, make a shot longer, change 16:9 to 9:16, drop a new scene into the middle of a clip, make a still photo talk, or orbit the camera around a picture.

They are partner (API) nodes, so they run on MiniMax’s servers and cost credits. They are not the open H3 weights you run on your own GPU.

This guide covers what each node does, the exact limits, the reported prices, how to set them up, a workflow for each one, and fixes for the common errors.

New to H3? Our MiniMax H3 ComfyUI guide lists every H3 workflow, file and fix in one place.

What the H3 Max Editing Nodes Are

MiniMax H3 is a video model with sound. There are two very different ways to use it in ComfyUI:

• The open H3 weights, which you download and run on your own GPU. Most of our H3 guides cover this route.
• H3 Max, MiniMax’s top closed model, which you reach through Comfy’s partner nodes. You pay per second, and nothing runs on your card.

The six new nodes belong to the second group. They were added by Comfy’s partner-node team in PR 17424 and appear in the v0.44.0 changelog as “MiniMax H3 Max editing nodes: Recast, Extend, Reframe, Insert, Lip Sync, and Camera Control”.

The same change also taught the existing H3 Reference to Video node to read @Image, @Video and @Audio tags in prompts, which is covered below.

All Six Nodes at a Glance

The limits below come from the node definitions as reported when the PR merged, and from the official Lip Sync node page. Prices are the reported per-second rates; your node’s price badge is the final word.

Setup

1. Update ComfyUI to v0.44.0 or newer. On a git install, run git pull in the ComfyUI folder and then pip install -r requirements.txt. On the portable build, run the update script in the update folder. Desktop users get it through the app’s updater. Our ComfyUI install guide covers each build.
2. Restart ComfyUI and refresh the browser.
3. Sign in to your Comfy account from the ComfyUI menu, or set a Comfy API key. Partner nodes do not run without one.
4. Add credits to the account.
5. Double-click the canvas, search for MiniMax H3 Max, and add the node you need.
6. Connect a Load Video or Load Image node, plus Load Audio where needed, and send the output to Save Video.

There is nothing to download into ComfyUI/models/, because the model runs remotely. Your input files go in ComfyUI/input/ as usual, and results land in ComfyUI/output/.

Recast: Swap the Cast

Recast keeps the motion, camera work, cuts and sound of a clip, and replaces the people in it with the people in your reference photos. It is meant for things like re-casting a storyboard clip with your own characters.

• Source video: 5 to 30 seconds long, and no single shot may be longer than 15 seconds.
• References: 1 to 4 photos, each at least 256×256.
• Order: by default, people are replaced left to right, so the first photo maps to the left-most person.
• Prompt: optional. Use tags like @Image2 to say which reference goes where, for example “@Image1 is the woman at the counter, @Image2 is the man by the door”.
• Resolution: 768P or 1080P, with an “enhance realism” toggle.

The docs work for this release adds a Cast Swap template built on Recast, and a reviewer noted that Recast rejects source clips outside an aspect ratio range of about 0.4 to 2.5.

Please use it responsibly. Only recast with photos of yourself, people who have given you clear permission, or made-up characters you generated. Swapping a real person into footage without consent can be illegal where you live, and it breaks the terms of most services.

If you want a local, open-weights way to reshoot footage instead, see our Viggle Remake guide.

Extend Video: Make a Shot Longer

Extend adds new footage after the end of a clip, guided by your prompt, with up to three optional audio references.

• Source: 1.625 to 60 seconds, up to 50 MB.
• New footage: 1 to 15 seconds at 480P, 768P, 1080P or 2K.
• Models: H3 Max or H3 Max Turbo. Turbo is reported at half the price.
• Output modes: extended returns the whole video with the new section added. continuation returns only the new footage, with a 1.6-second overlap so you can blend it yourself.

Tip: describe what happens next, not the whole scene again. “She turns and walks out through the door, the camera stays still” works better than repeating the original prompt.

Use continuation if you edit in a video editor, because a short clip with an overlap is easier to cross-fade. For a free local way to chain clips, see our H3 Extender multi-clip workflow.

Reframe Video: Change the Aspect Ratio

Reframe turns a landscape clip into a vertical, square or ultra-wide one without cropping. It extends the canvas and generates the missing edges, and keeps the timing and sound.

• Target ratios from a 16:9 source: 9:16, 1:1, 4:3, 3:4, 21:9 and 9:21.
• Source length: 0.5 to 300 seconds.
• Resolution: 480P, 768P or 1080P.
• Prompt: optional, to describe the new surroundings, for example “a busy street with shop signs above”.

This is the node most YouTube and Shorts creators will use. One horizontal video can become a vertical Short without losing the subject off the side of the frame. The docs PR adds an Aspect Ratio Reframe template for it.

Insert Video: Add a New Scene Mid-Clip

Insert replaces part of a clip with a new scene and then cuts back to the original.

• Timing: set start_time (where the new scene begins) and resume_time (where the original picks up again).
• New scene length: 5 to 13 seconds.
• References: up to nine images (@Image1 to @Image9) and up to three videos (@Video1 to @Video3, 15 seconds in total).
• Options: colour match to the source, and 480P or 768P.

The node’s own tooltip warns that references without a prompt are often ignored. So always describe the new scene in words and mention the tags, for example “cut to a close-up of @Image1’s hands typing on the laptop”.

For local frame-level edits on open H3, our H3 section editor guide and H3 Ripple guide are the closest free options.

Lip Sync: Make a Photo Talk

Lip Sync animates a still image so the mouth follows your audio. The video is as long as the audio. These are the inputs from the official node page:

The output aspect ratio is the supported ratio closest to your image. Turn on transcription for speech, because it helps the mouth match the words. 899 seconds is just under 15 minutes, so a full narration can run in one go.

The same consent rule applies: use your own face, a person who agreed, or a made-up character. For a voice track, our guides to Fish Audio Drama 3 and other local voice models can make the audio for free.

Camera Control: Orbit Around a Still

Camera Control makes a short clip where the camera moves to a new angle and distance around the subject in one image.

• Azimuth: -360 to 360 degrees (left and right orbit).
• Elevation: -90 to 90 degrees (up and down).
• Distance: 0.1 to 10, where 1 keeps the current framing, lower moves closer and higher moves away.
• Length: 1 to 15 seconds at 480P, 768P or 1080P.
• Helpers: it can be driven by a Create Camera Info or Compose Camera Angle Prompt node.

Two tips from the node notes: an empty prompt freezes the scene, so only the camera moves, which is ideal for product spins. And azimuth is a target, not a promise, so a 90-degree orbit may land anywhere between about 30 and 90 degrees. Run two seeds if the angle matters.

The @Image, @Video and @Audio Tags

The same PR updated the existing H3 Reference to Video node so prompts can point at connected inputs with short tags:

• @Image1, @Image2 and so on refer to the image inputs in order.
• @Video1 and @Audio1 work the same way for video and audio inputs.

The node resolves each tag to the matching reference before sending the request. Only use tags for inputs you have actually connected. A tag pointing at an empty slot can stop the run with an error, depending on the node.

For the local open-weights version of reference video, see our H3 Ref2VA guide.

Keeping the Cost Down

Every node is billed by the second, and costs rise quickly with resolution. A few habits keep the bill small:

• Test at 480P, then re-run the version you like at a higher resolution.
• Trim inputs first. Cut the source clip to just the part you need before Recast or Reframe.
• Use Turbo for Extend drafts, since it is reported at half price.
• Use continuation mode on Extend if you only need the new part.
• Read the price badge on the node before you queue, because pricing can change.
• Fix the seed while you adjust the prompt, so you compare like with like.

API Nodes vs Local Open H3

A sensible split is to use local H3 for drafts and experiments, and the paid nodes for the one or two edits local tools cannot do yet, such as Reframe. If your card struggles with local H3, our H3 out of memory guide helps.

Troubleshooting

I cannot find the H3 Max nodes

Update to v0.44.0 or newer and update the frontend too. Then make sure you are signed in, because partner nodes can stay hidden or fail without an account. Some nodes may also be gated by region or plan.

Unauthorized or “please log in” error

Sign in again from the ComfyUI menu, or check that your Comfy API key is set. If you run ComfyUI headless, the key has to be passed to the server.

Insufficient credits

Top up the Comfy account. The price badge shows the cost before you run, so lower the resolution or length if a run costs more than you expected.

Video rejected for length or size

Check the limits table above. Recast needs 5 to 30 seconds with no shot over 15 seconds. Extend needs 1.625 to 60 seconds and under 50 MB. Trim or compress the clip first.

Image or video rejected for aspect ratio

Lip Sync needs an image between 0.4 and 2.5 in aspect ratio, and at least 256×256. Recast has a similar range for source clips. Crop very tall or very wide inputs.

Insert or Recast ignores my reference

Mention the reference in the prompt with its tag, such as @Image1, and describe what it should do. References without a prompt are often ignored.

Camera Control did not reach my angle

The azimuth is a target. Try another seed, or ask for a slightly larger angle than you need.

FAQ

Are the H3 Max nodes free?

No. They are partner nodes billed per second through your Comfy account.

Do I need a powerful GPU?

No. The model runs on MiniMax’s servers, so even a laptop can run these nodes.

Can I use them with the open H3 model?

They are separate. You can feed a clip made locally with open H3 into an H3 Max edit node, because the node only needs a video file.

Which node is best for YouTube Shorts?

Reframe Video, because it turns a 16:9 clip into 9:16 by generating the missing top and bottom instead of cropping.

How long can a Lip Sync video be?

Up to 899 seconds of audio, just under 15 minutes. The video matches the audio length.

Is it legal to recast or lip-sync a real person?

Only with that person’s permission. Laws on synthetic media differ by country, and using someone’s likeness without consent can be illegal. Stick to yourself, willing participants or made-up characters.

Sources

• ComfyUI changelog (v0.44.0 entry)
• ComfyUI PR 17424 (H3 Max nodes and @tags)
• ComfyUI docs: H3 Max Lip Sync node
• Comfy docs PR 2243 (Cast Swap and Aspect Ratio Reframe templates)
• ArtRealmAI: ComfyUI adds six H3 Max nodes (limits and reported prices)