MMH3 Temporal Tile Editor

MMH3 Temporal Extend Video does the sampling, but you don't configure a multi-segment H3 clip by clicking sockets on it. You do it here - in a dock panel that opens off this node, where each segment gets its own prompt, seed, references and duration, and the whole thing is serialized into a config the Extend node reads.

It's a pure-information node, and that's worth internalising. It loads no images and encodes nothing. Any audio you wire into it is forwarded verbatim to the Extend node, which owns the audio VAE. Think of it as the plan; the Extend node is the execution.

What you set, per segment

Segment 0 is special: it starts a fresh generation, either FL2VA (first frame, optional last frame) or Ref2VA (reference images). It also fixes the chain's resolution - those width/height values get rounded to multiples of 32 and are locked for every later segment once a session has stored latents. Choose them before you commit to a long run.

Later segments continue the previous tail and have one continuation knob of their own: overlap frames, the length of the carried tail. Everything else about continuation - fade, anchors, blending - is global, set on one of the two overlap nodes. That split is deliberate and it's the thing to remember when tuning: per-segment length, chain-wide behaviour.

The rest of a segment row is what you'd expect: prompt and negative, a seed (there's a randomize-per-segment option), conditioning mode (Ref2VA, or FL2VA with only a last-frame keyframe - frame 0 belongs to the seam anchor), a reference source (manual, prev_frame to take a frame of the previous segment, or none), reference image sizing (match or max, aspect preserved), and an audio reference mode: none, load (a picked file), prev (the previous segment's soundtrack), initial (the first segment's), or bgm.

Durations are grid-bound and the panel says so: the first segment must be 17n+5 frames (5, 22, 39, 56, 73, 90, 107, 124 - 124 ≈ 5.2s is the default), and later segments generate a multiple of 17 new frames (102 ≈ 4.25s by default), because H3's VAE codes 17 pixel frames per 5 latent tokens. Do the arithmetic in the panel rather than in a separate node - the panel exposes both frames and seconds.

The sockets

The required widgets are show_editor, session_name, storage_location, base_prompt, base_negative and tile_data. Two of those deserve a word. base_prompt and base_negative are shared prefixes applied to every segment, which saves you repeating a style line six times. And tile_data is the JSON the dock panel owns - if you edit it by hand, expect the panel to rewrite or sanitize what you typed.

Optional inputs: ref_images, an autogrow set of IMAGE sockets (First_or_Ref_Image_0, Last_or_Ref_Image_1, then Ref_Image_2…, up to 16), and audio_bgm, a single song for the whole chain.

audio_bgm is the more interesting one. Wire a track roughly as long as your finished video and set a segment's audio reference mode to bgm: that segment's soundtrack is cut from the song at its own position on the chain's timeline, encoded through the Extend node's audio VAE, and frozen by the audio noise-mask. The music is the audio; the model generates video on top of it while still hearing it. Encoding happens one segment at a time, so peak VRAM tracks a single segment rather than the whole song. If the music runs out mid-chain, the mask drops and the model continues it freely.

The three outputs

temporal_tile_config goes to the Extend node. segments_info is an inspectable dict - session, storage, segment count, per-segment parameters - genuinely useful for confirming what the panel actually meant. ref_image_slots carries the wired sockets to the Extend node's matching input as one record per socket, at native resolution.

That last detail is the source of a common confusion: images are never resized or merged here. One socket is one <Picture i>, and a socket fed by a batch contributes only its first image - the same convention as MiniMaxH3 Reference to Video. Resizing happens per segment, in the Extend node, according to that segment's reference image size. The row order in the panel is the prompt order, so if your prompt says <Picture 2> and the result ignores you, check the chip order first.

Sessions, previews and re-rolls

Everything about a run lives in one session folder, <temp|output/latents>/mmh3_temporal/<session_name>/ - a session.json ledger plus latents/ and previews/. Segments can be locked, meaning they're read back from disk and never re-sampled. Unlock one segment, change its seed or prompt, run again: only that segment re-samples, and its previous attempt is preserved. The panel's previews let you click and set resume_from_segment, and editor state is stored in the session, so reloading the workflow restores the panel.

Two consequences. storage_location: temp is wiped when ComfyUI restarts - fine for experimenting, wrong for a session you intend to return to; use output/latents for anything long-lived. And disk fills up quietly, which is what the panel's "clear unused latents" action is for. There's also a shared reference/conditioning cache beside the sessions, content-addressed on the model fingerprint, the source file's path/mtime/size or a pixel hash, and the encoding parameters - so references and prompts are encoded once across segments, runs and sessions, and change an input and it simply misses.

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 pip step. If the dock panel doesn't appear after installing, reload the page - node panels only load on a fresh front-end.

Two more expectations to set: the pack is at version 0.0.2 and its README warns that sockets, parameters and on-disk formats may still change between commits. And if a segment uses bgm without an audio VAE connected on the Extend node, the run stops before sampling with an explicit error - it fails loudly instead of shipping a silent clip.