MMH3 Temporal Extend Video

MiniMax H3 gives you 4-15 seconds. Everyone eventually wants a minute. The standard workaround, on every video model going back to AnimateDiff, is last-frame chaining: decode clip one, grab its final frame, feed it in as the first frame of clip two, repeat. It works and it drifts - you re-encode a decoded frame every round, the model re-invents details it already had, and the seam is a cut no matter how careful you are. That's the pain this node attacks.

This is the main event of the pack. Everything else - the tile editor, the sampler bundle, the overlap knobs - exists to feed it.

What it does differently

Instead of chaining in pixel space, it chains in latent space. A segment's sampling latent starts as a copy of the previous segment's tail, snapped to a video-token boundary on H3's keyframe grid, and the appended part is sampled freely. The carried tail isn't noise-conditioned or re-encoded; it's held in place by the latent noise mask - frozen zone, then a fade band, then free generation.

The grid arithmetic is the thing that trips people. H3's VAE codes 17 pixel frames per 5 latent tokens, so valid H3 lengths land on 17n+5: 5, 22, 39, 56, 73, 90, 107, 124. At 24 fps the default first segment of 124 frames is ≈ 5.2s, and continuations add 102 new frames ≈ 4.25s. It's the same class of trap as Wan's 4n+1 rule with a different constant, and the pack snaps values for you rather than letting you fail at sampling time.

Within a segment, the mask value becomes each token's timestep label and mixes that token's input, so a constant band of mask reads as "partially preserved content". A frozen zone, a gradient band and free tokens then coexist in one sampling pass, and a seam anchor can pin the conditioning at the split frame so the join doesn't wander.

The inputs that matter

clip (the H3 CLIP that builds each segment's conditioning), vae (the H3 video VAE that encodes reference images and keyframes), temporal_tile_config from MMH3 Temporal Tile Editor, and sample_params from MMH3 Sample Params. That's the spine.

Optional, in rough order of how often you'll want them: overlap_params (either overlap node), audio_vae (needed only when a segment references an audio file; latent-level references like "previous audio" don't need it), 2nd_sample_params (a refinement pass that re-denoises a whole continuation segment at low denoise with no mask and no keyframes), latent_seg_0 (extend an H3 latent you already have instead of starting fresh), and ref_image_slots (the tile editor's wired image sockets). second_pass_audio picks whether the stitched soundtrack comes from the second pass or the untouched first pass.

Outputs and where they go

merged_latent is the accumulated timeline - this is your video. segment_latent is the last segment standing alone, which is what you want when you're inspecting just the piece that ran. segment_info is a summary dict: session directory, files written, split points, sampling stages.

H3 latents are audio-visual pairs (video and audio in one nested tensor), so decode both: video VAE into VAEDecode, audio VAE into VAEDecodeAudio, then CreateVideo at 24 fps and SaveVideo.

Every finished segment is also written into the session directory as a .h3latent plus a preview WebP, which is what makes resume and re-rolls possible: lock the segments you like, unlock the one you don't, and only that one re-samples. Earlier attempts are kept.

Install

ComfyUI Manager → search Comfyui-MMH3-UltimateExtend → install → restart, or:

cd ComfyUI/custom_nodes
git clone https://github.com/bbaudio-2025/Comfyui-MMH3-UltimateExtend

No dependency install - the pack leans on ComfyUI's own torch/PIL and the H3 support in core. What you need on disk is the H3 stack: diffusion model, CLIP, and the video and audio VAEs. The H3 weights ship under a community licence that voids the grant in the US, EU, UK and South Korea, outputs included.

Two caveats before you invest time. The README carries an explicit work-in-progress warning - sockets, parameters and on-disk formats may change between commits - and this is version 0.0.2. And the spatial half of the pack (MMH3 Spatial Extend Video) is still under development; the temporal chain is the usable part.

Common problems

The errors here are unusually explicit, because they're mostly raised by the node itself:

• FL2VA needs a VAE - a segment is using a first/last frame as a keyframe but no video VAE is connected. Without it the frame would only reach the text encoder and the output would quietly ignore it, so the node refuses rather than lying to you.
• the 'bgm' audio reference mode needs a VAE or reads the song from the 'audio_BGM' input ... which is not connected - a segment is set to take its soundtrack from the audio_BGM socket of the tile editor. Either the song isn't wired or audio_vae is missing.
• resume_from_segment=N but no stored merged latent found - you're resuming from a session with no saved state. Run the chain once. Related: if storage_location is temp, those files are cleared when ComfyUI restarts, so a session you meant to come back to is gone. Use output/latents for anything long-lived.
• fade_frames=... exceeds the realized tail length - your fade band is longer than the tail you actually carried. Lower the fade or raise the segment's overlap frames in the editor.
• latent_seg_0 must be a MiniMax H3 AV latent - you fed it a plain video latent or an image latent. It wants an H3 audio-visual latent, batch size 1.
• Import fails outright - an older ComfyUI that predates MiniMax H3 support and the comfy_api.latest schema. Update core.
• You hear "it's slow" - that's H3, not this node. Chaining multiplies full sampling passes, and every segment gets a decode on the end.