MiniMax H3 0.6 example workflows

These are clean ComfyUI UI-workflow documents built for the 0.6 node surface. They use fresh node serialization, the organized Production Plan, explicit Generation Profile, and numbered, non-overlapping project/generation columns. Older 0.5 files are preserved unchanged in Archive/0.5/.

For installation and first-run setup, see Getting started.

Each workflow has a compact START HERE note. The longer setup and wiring instructions are in its matching Markdown file under guides/, so they no longer take up large empty canvas panels. Titles and preview sizes were checked in ComfyUI, including the VHS preview extension.

The catalog is compiled from named recipes, not archived workflow JSON. All H3 sockets, widget positions, and values are checked against the 0.6 checkout. Model files, source tracks, and extra packs still need to be installed/selected as described below; schema validation does not replace a GPU render test.

The loader defaults use the canonical filenames from the official Comfy-Org MiniMax H3 package: put diffusion weights in models/diffusion_models/, the Qwen3-VL encoder in models/text_encoders/, and both VAEs in models/vae/. The workflows do not assume a custom model subfolder.

Choose a starting point

Input or task	Recommended workflow
Text only	T2V Normal
Text-only timeline, takes, and trims	T2V Studio
One opening image	I2V Normal
Image-led timeline, takes, and trims	I2V Studio
First and last images	FL2V Normal
A few global Ref2VA pictures	Ref2V Basic
Prompt-selected `@tag` references	Ref2V Tagged
Tagged references with full project authoring	Ref2V Studio
Project references plus exact source audio	Ref2V Studio Source Audio
Video inpainting	Masked Video Inpaint
Ref2VA-guided video inpainting	Ref2V Masked Video Inpaint
Continue existing video	Masked AV Extension — Single Clip
Extend existing video over several scenes	Masked AV Extension — Chain
Generate the exact gap between two clips	Masked AV Bridge — Two Clips

Normal and Studio

Normal workflows keep the familiar scene-column Production Plan and the standard prompt editor. They are the smallest useful graphs.

Studio workflows add the Plan Studio timeline, Project Asset Carousel, rich prompt editor, and Checkpoint Manager. The Carousel owns the project/run name, reference lineage, and optional Source Track. Checkpoint Manager browses saved scenes, alternate branches, trims, previews, and restoration state. These authoring tools do not change the sampler body.

To initialize a Studio reference project:

Copy or import the two 0.6 courier images from assets/.
In Project Asset Carousel, tag the arrival image courier_arrival and the delivery image greenhouse_delivery; assign both the Picture role.
In the Source Audio workflow, also import one audio file and assign Source track. Generation Profile is already set to Lip-sync to source audio.
Edit the example scenes and queue the workflow.

For vocal-only lip-sync, import aligned full-length stems and assign them under the Source track's Synchronized audio tracks. The full mix remains the soundtrack; vocals guide generation. Use the scene Lip-sync selector in Plan Studio for instrumental/action-only scenes. See grouped audio.

The Carousel stores media as ordinary project assets rather than embedding image bytes or stale file bindings into workflow JSON.

Current 0.6 authoring contract

Every maintained recursive workflow uses:

MiniMax H3 Generation Profile for scene continuity and audio intent;
MiniMax H3 Production Plan for scene columns, timing, seeds, and output;
20-step res_multistep / simple sampling defaults;
explicit carried-overlap language in continuation prompts, so new action begins after the inherited boundary rather than being cut short;
a muted manifest/assembly recovery path where appropriate.

In I2V workflows, do not bypass Frame Gate. It prevents the opening picture from being reapplied to every scene. FL2V uses Frame Index Switch plus Frame Gate to demonstrate an A→B→A endpoint sequence.

For prompt syntax, see Scene authoring. For tags, motion references, and source media, see Scheduled references.

Masked editing and existing video

The masked workflows keep the bundled CC0 soldier-crab source because it gives the mask, source-video, source-audio, and protected-boundary examples a legal, reproducible common input. Copy the required files from assets/ to ComfyUI/input/ before opening those workflows.

The two-clip bridge is a single masked target with an ordinary sampler. The extension and inpaint workflows use the checkpoint/review/resume loop. See Masked editing for H3 grid behavior and audio protection.

Deferred upscale

Deferred workflows start from a lineage selected in Checkpoint Manager; they are not first-install tests.

Workflow	Extra requirement
SeedVR2 Full Chain	ethanfel SeedVR2 fork
H3 LBH 3D	LBH H3 latent upscaler
H3 LBH 3D + De-Rope	LBH pack plus ComfyUI-MAINodes
Pixel DLSS5 + USDU — Experimental	DLSS5, H3 USDU Guider fork, and Turbo v4 LoRA; setup and testing limits

Upscaled variants are saved below output/h3_chains/<run>/upscaled/<profile>/ and never replace source checkpoints. See Runs and recovery.

The SeedVR2 workflow uses the current SeedVR2 Video Path Upscaler node (SeedVR2VideoPathUpscaler), not the retired SeedVR2DirectVideoUpscaler ID. Install the base SeedVR2 model-loader pack as well as the linked video-path extension. For LBH, the temporal-chunking and unload controls are explicitly saved; the De-Rope injection preset is custom, with 20 total steps and 0.5 injection, so it does not silently override the visible schedule.

Assets and provenance

The arrival/delivery pictures were generated specifically as neutral 0.6 reference assets. The masked examples use CC0 footage. Exact hashes, licenses, copy instructions, and the older 0.5 reference provenance are documented in assets/README.md.

ComfyUI cannot load arbitrary files from the custom-node repository directly; copy them to ComfyUI/input/ or import them through Project Asset Carousel.