H3ZeroStudio brings a video-editor-style timeline into ComfyUI for MiniMax H3. Instead of treating every generation as an unrelated graph run, it organizes segments, references, alternate takes and sound inside a project. Its central distinction is between continuing a previous segment and starting a new scene.
This guide follows the H3ZeroStudio repository, including its installation, project-storage and audio sections, reviewed October 22, 2026. The steps describe the author's implementation; we have not installed the pack or verified its continuity claims. H3Vid's online generator does not currently import H3ZeroStudio projects.
Install the editor around a working H3 setup
Use a recent ComfyUI build with native MiniMax H3 support. In ComfyUI Manager, search for H3ZeroStudio, install it and restart. The repository also documents comfy node install h3zerostudio, or a manual checkout:
cd ComfyUI/custom_nodes
git clone https://github.com/zeroleloushe/H3ZeroStudio.git ComfyUI-H3ZeroStudio
The pack uses dependencies already supplied by ComfyUI for its normal operation. Optional RTX video processing has additional requirements, so it should not be part of your first basic run. After restarting, find H3ZeroStudio · Timeline in the MiniMax H3/H3ZeroStudio category.
Connect model, clip, vae and audio_vae. A missing audio VAE leaves you with video-only operation. The editor does not replace the H3 transformer or provide model weights. Establish a working H3 generation graph before debugging a long project.
Build a short project before a long film
Expand the Timeline node, upload the character, location and sound assets, and attach the relevant inputs to a segment. Write each segment's action and choose its relationship to the previous one. Use Continuation when it should extend the previous state; use New scene for an editorial cut.
The author describes continuation as retaining the previous latent tail, using the last frame as a reference and continuing sound through H3's audio mask. That mechanism can help continuity, but it does not guarantee a character will survive arbitrary pose, lighting or scene changes unchanged.
Begin with two short segments featuring one subject and a clear handoff. For example, a traveler approaches a bench in the first segment and sits in the second. Inspect posture and location at the junction before adding a third segment. A deliberate cut to a new camera angle should be a new scene, rather than a forced seamless continuation.
Reference labels belong to each segment
Assets can be grouped as characters, locations, other images and audio. The editor numbers image references automatically, with up to nine pictures per segment. Its prompt editor offers reference insertion and highlights labels that do not resolve.
Check the segment's actual input order before writing about Picture 1. Reusing a familiar filename elsewhere in the project does not establish its number in the current prompt. Keep the action description specific enough that each reference has a role: appearance, setting, clothing or motion.
Character cards use RefMod safetensors and a shared library. They are different from Omnichar character files. Project export does not imply those two formats are interchangeable, and a shared character library deserves its own backup when moving the workspace.
Takes, cache and dependent segments
H3ZeroStudio keeps up to twelve takes per segment. The chosen take feeds the assembled film and the next continuation. Changing it can make downstream segments stale, because those segments were generated from an earlier state.
Use a new take to explore a different gesture while retaining the approved version. After choosing it, review dependent continuations before exporting. A cached output is useful only if its recipe and upstream state still match the project you intend to deliver.
The README describes content-based reference checks, separate draft and final caches, and a restore view that explains changed fields. Project assets, outputs and caches live under output/h3zerostudio/<project-id>/; project metadata also exists inside the workflow, with copies under _projects. A workflow JSON and a complete ZIP with assets serve different purposes. Test an exported project's contents before deleting its original folder.
Locked audio and background music do different jobs
The Audio lanes fix supplied sound in the audio latent while video adapts to it. The Music lane mixes a soundtrack during assembly without conditioning the video on that music. An audio asset attached as a segment reference is a third case: it guides generation rather than promising exact waveform preservation.
Choose the Audio lane for approved speech or a musical performance that must drive timing. Choose Music when the clip is already approved and only needs a background bed. Listen across continuations and scene boundaries; a clean picture junction does not prove the dialogue has no doubled syllable or level jump.
Refinement and partial regeneration
The two-stage mode makes a smaller draft, enlarges its selected video latent and runs a refinement pass while preserving draft sound. The documented starting sigma is 0.40; this is a noise level, not a universal denoise percentage. Learned latent upscaling requires its separately listed model and node.
Partial regeneration uses context around a selected region and protects the surrounding video and audio through masks. The README's current limitations matter: a repair cannot cross an ordinary new-scene cut or certain other discontinuities. Segment boundaries follow H3's frame grid, so a selection may not match an arbitrary millisecond edit.
The node outputs images at 24 fps, audio and an assembled AV latent. Connect the appropriate downstream save route and inspect the resulting file. For a first project, complete two segments, save the workflow and assets, reopen them, then export. That exercises the practical project handoff before a lengthy queue commits you to the setup.
The cover is an imagegen concept illustration of timeline editing, not an H3ZeroStudio screenshot or H3 output.