Long MiniMax H3 videos can be built through several short generations, but those workflows carry state differently. H3-LongVideos organizes written beats into shots. ComfyUI MiniMax H3 Timeline Director exposes editable media assignments and segmented continuation. Both deserve a closer look than the phrase “unlimited video.”
This guide reviews current author documentation and the LongVideos node source on October 17, 2026. It does not report an H3Vid long-render test. The cover is a generated concept illustration of joins and story beats.
Choose by the project you need to control
The H3-LongVideos repository describes a chain of shots, up to roughly 120 seconds, using ComfyUI's native H3 support. Its first paragraph supplies a persistent scene anchor; subsequent paragraphs provide beats. Character memory and optional reference images help specify who appears and how they look.
Timeline Director instead gives an editable media timeline and segment plan. Its all-in-one workflow can organize reference material, per-segment directions and continuous generation. It also supports a direct latent continuation path and an optional progressive-resolution sampling route.
Choose LongVideos when prose beats are the useful planning unit. Choose Timeline Director when explicit segment boundaries and media assignments are central. Neither choice removes the need to review a completed shot. A longer sequence creates more opportunities for identity, spatial arrangement or dialogue to drift.
Build a small LongVideos graph first
The documented graph connects the H3 model, text encoder and VAE loaders into H3 Long Videos, then sends images and audio to a video-combine node. The latent output is optional post-processing data. Required base components remain separate from the custom-node download.
Copy the node folder into ComfyUI/custom_nodes and restart the server. The current documentation requires ComfyUI 0.31 or later with native H3 support. Use the actual node file to confirm the revision you installed rather than assuming every screenshot describes the same fields.
Text fields such as prompt are sockets expecting a connected multiline text node. A missing prompt connection is not an empty text box to type into. Shot length also has a dedicated connection. These details matter when a graph appears complete but cannot validate.
An illustrative two-beat plan is:
A quiet seaside station at dawn. One traveller wears a navy coat and carries a yellow canvas bag. The camera remains at platform height.
The traveller walks toward the bench and places the bag beside it.
The traveller sits and watches the train approach. No dialogue.
Keep the action modest enough to inspect one boundary. If you add dialogue later, assign each line clearly and give it sufficient shot time.
Preview the plan and inspect output semantics
Use plan_only to inspect the proposed structure and info before a GPU render. This can expose unexpected size, reference assignment or compatibility warnings. It does not prove the final motion or speech will work, but it catches planning mistakes cheaply.
The critical LongVideos distinction is between decoded images and joined sampled latents. The chain needs decoded frames for subsequent handoffs. Seam trimming and handoff offsets operate on frames, while H3 compresses time in latent space. Consequently, the multi-shot latent output is not an exact latent representation of the final trimmed images. A single-shot run avoids that particular mismatch.
Do not wire a latent post-processing result into an editing pipeline and assume its timing matches the finished frame sequence. First identify which output owns the intended delivery timeline. The repository's reference document is explicitly marked out of date by the current README; use it for background and check active fields against source and tooltips.
Set up Timeline Director from its own workflow
Start from the all-in-one examples, not by transplanting LongVideos socket assumptions. Timeline Director carries prior AV-latent context, applies aligned overlap handling and assembles output. The author's two approximately one-minute examples are 1,263 frames at 24 fps, or 52.625 seconds. They are selected creator demonstrations, not an unrestricted hardware promise.
Its changelog also describes an unreleased low-memory sampler that writes temporary encoded segments and exposes a VIDEO output directly to SaveVideo. That output needs different wiring from an image-batch path. It is not a saved-latent checkpoint system, so temporary spilling should not be mistaken for restartable generation state.
Progressive-resolution sampling is another independent choice. The documented eight-step example allocates six steps at low resolution and two after a latent lift. It requires a compatible H3 latent upscaler; an LTX checkpoint in the same folder is not equivalent. Establish ordinary sampling before adding this extra stage.
Preserve the intended soundtrack and verify joins
A source recording retained during final export is different from generated audio continuity. Timeline Director's locked-original-audio route and LongVideos' soundscape or silence controls should be read within their own contracts. Do not transfer a switch's meaning between node packs merely because both mention audio.
After rendering, play every boundary with sound. Inspect duplicated poses, costume changes, camera jumps, mouth motion and soundtrack discontinuities. Retain the graph, seed, prompts, source media and revisions with the result. If a later change fixes one join but breaks another, those records make the comparison useful.
H3Vid's online Extender offers a separate finished-video workflow. It does not expose these local planners, latent handoffs or memory controls. For local work, prove two segments first, then increase duration with a clear plan for memory, disk use and recovery.