ComfyUI_zura_nodes

Give the video model a voice track it can actually hear

The two-field handshake that stops you paying twice

When SAM finds four people and you only wanted one

Fake-colour depth maps, because Seedance can't read a mask

The boring arithmetic that makes MiniMax H3 accept your clip

Turn a grey blockout into something H3 will render from

Reload H3's character references without re-encoding them

The approval gate that re-hashes your source before rendering

Pick up a cached video prompt in a fresh job

Why your LTX clip has to be 8n+1 frames long

A receipt for the paid draft you just generated

Keep the original audio, no matter what the model did to it

Cache the expensive half of a MiniMax H3 take

The node that makes you look at the mask before it renders

Encode the prompt once, sample it all night

Cut your source clip to 24 fps before anything else touches it

Making the model repaint the captions too

Trust the file, not the graph settings

Pitch-shift the voice so the model stops copying it

The 24 fps clock at the front of a Wan 2.2 Animate run

Zura Mask takes any video, not just Wan's — and that's the useful part

Zura's Looped Chunks Sampler vs Wan's five-second wall

Two model branches, one checkbox, no re-wiring

Plan Nine Camera Angles, Then Shoot the Clip You Planned

The same audio swap, but paranoid and memory-bounded

Gluing two MiniMax H3 windows together without a visible seam

The Node That Stitches Your Multicam Shots Back Together

The one-node version of Zura's dialogue editor — and when to skip it

Change the selection, keep the voice take

Zura's replacement-words editor

Get the reference and the timings right

The three-wire node that makes Zura's auto word selection work

Turn the Bedroom Take Into a Studio Session

Paste a mouth back into the shot without retiming a single frame

Zura Select Words and Face

The five-second fix for a mask full of pinholes

Wiring MiniMax H3's ControlNet the way the sampler expects

Hand-encoding the character H3 should keep

Stop the model rebuilding subtitles on your actor

Make H3 use the take you actually chose

Tired of H3 drifting off your approved first frame? Pin it instead

Multicam Inside MiniMax H3

Generative Multicam That Moves the Set

Why your control payload is 49 channels

Getting your source, depth and mask onto H3's grid

Kill the per-window audio drift before it happens

Keep the room, change the actor

Your performance, from the other side of the room

Decoding one slice of a long clip without decoding the whole thing

Grey Out the Room, Keep the Person

Stitching per-shot masks back together, and telling you where tracking failed

Nine Widgets, One Identity-Safe Relight Prompt

Reading back the previous window, with proof it's the right one

Split a huge graph in two and stop re-encoding for free

Same LongCat voice, except it stops re-synthesising every time you nudge a handle

Local voice cloning that runs in its own little Python bunker

The tiny converter between where you clicked and what SAM wants

Fix the Guide's Mouth Before the Model Ever Sees It

One Dropdown That Keeps Your VRAM Alive

A/B Switch With an Off Position That Actually Saves Time

Prune the Planning Branch Instead of Muting It

Invent a voice by describing it, no recording needed

Camera-Angle Prompts That Don't Decapitate Your Subject

Reload a Planned Camera Angle Without Regenerating It

The two tensors that make a seamless H3 join possible

Park an encoded video latent on disk and sample it later

Make the Room Turn With the Camera, Not Just the Performer

The boring node that stops your take landing one frame off

Paste the animated mouth back and walk away with a VIDEO

Animating a face crop before you render anything

Pick your mouth engine before you pick a fight with the install

Stop your tracker from chasing one person across two shots

Pulling one value back out of a stage bundle

Getting graph-view audio back into the artist UI

A poor man's context bus that costs you type safety

Deleting on-screen text without touching your actor

The 41-frame loop that makes Wan Animate hold a face for 30 seconds

The same Wan Animate loop, with the frames going straight to disk

Why your actor's legs vanish on section two

The node that tells you your graph is stale before it renders

The mmap loader for very long shots

The cached footage payload, in one wire

Reading a 30-second shot 81 frames at a time

Stitching the per-section shards into one cache

The six wires the sampler really needs

Decoding one 81-frame slice instead of the whole take

What Studio actually approved, as tensors

The node that freezes your approved take before Wan eats 25 minutes and all your VRAM

One section at a time

The node that tells you your face motion and your audio don't match — before the 25-minute render

Not a face-crop node — a bouncer that checks your performer's ID first

How Wan's native speech guide gets recorded

The most boring node in the pack, and the one you want for a whole-scene edit