MMH3 Temporal Overlap Params

When you chain H3 segments, the interesting part is the join. The previous segment's tail gets carried into the new segment and held there by a latent noise mask - a frozen zone that keeps what the model already generated, a fade band where the new content takes over, and then free sampling. This node is the control panel for that band. Its sibling, MMH3 Temporal Overlap Simple, is the same socket with four presets; come here when a specific seam looks wrong and you want to know why.

One thing it does not control: how long the carried tail is. That's per segment, set as overlap frames in the tile editor. Everything on this node applies to every continuation segment in the chain.

The knobs that actually matter

tail_mode - freeze_fade (the default) or free_resample. Freeze+fade pins the frozen zone and fades into the new content; free resample lets the entire carried tail re-sample, which sounds more elegant and is flagged experimental.

fade_frames - length of the fade band in pixel frames. 0 means no fade at all: the whole tail stays frozen and the new content hard-cuts in. Anything else snaps up to a multiple of 17, because that's H3's keyframe grid (17 frames per 5 latent tokens). It has to fit inside the carried tail, or sampling stops with fade_frames=... exceeds the realized tail length.

fade_value and fade_mode shape that band: how strongly the transition lands (0 = hard cut toward the new content, 1 = keep the tail) and what curve it follows across the width. Defaults of 0.5 and a flat curve are fine until you're chasing a specific look.

seam_reference - set to prev_tail and the previous merged video's tail is fed to every continuation segment as a bit-perfect latent video reference, soundtrack included, through H3's native Ref2VA mechanism. No soft mask is involved, so none of the mask artifacts below can happen. It's designed to be used with fade_frames=0 and a generous overlap, with seam_ref_frames deciding how much of the tail the reference covers (snapped to the 17m+5 grid: 5, 22, 39…; 0 = the whole carried tail). Highest-quality seam option, slowest, because the model attends to a video reference on every segment.

max_mask_strength - how strongly the mask does its job. 1 holds the frozen zone completely; 0 removes the mask entirely, so the whole tail re-samples (equivalent to free resample). If your chain is slowly bleaching or softening, this is the one-value experiment that tells you how much of it is the mask.

The knobs you reach for second

anchor_seam and anchor_strength pin the conditioning at the split frame so the join sits where you put it - 0.999 is a near-hard pin. keyframes_mode decides what happens to the previous segment's keyframes after the split: reanchor (move them to the new split point, the default), drop, or keep.

overlap_mode and overlap_blend decide who wins the overlap band - later (the new segment) or earlier (the accumulated video) - and how that transition is shaped. Worth knowing: when a fade band exists, the blend ramp is confined to the frozen zone, where both sides carry identical content, and the re-developed fade band is handed wholly to the new segment. Crossfading two genuinely different renditions of the same frames would ghost.

The experimental corner

fade_impl has two values and is flagged as possibly going away. mask is the default soft noise band - intermediate per-row timesteps - and it is known to produce mosaic-like artifacts on H3. qsample_init keeps the band at full-strength timesteps instead and steers the transition by blending the carried content into the initial noise, decaying as sampling proceeds. The author's claim is the interesting part: because the band never sees intermediate timesteps, subsequent segments continue far more reliably, at zero extra cost.

That comes with a caveat worth taking seriously. The structured init at sigma_max is out of distribution and the model over-develops it, which shows up as brightness and saturation drift in the band. init_content_weight (default 0.3, only effective with qsample_init) caps how much carried structure rides in that noise; 0 is pure noise. Negative values carry sign-flipped, anti-correlated structure - an experiment, not a setting. The author's own framing: the start stays pinned to the reference but inverted, so expect mirrored development or nothing at all.

The sampler you pick changes all of this. Both tooltips say the same thing: use a stochastic SDE sampler that re-injects fresh noise every step - sa_solver, er_sde, dpmpp_2m_sde, dpmpp_3m_sde. Deterministic ODE samplers (euler, res_multistep, uni_pc, dpmpp_2m) are more likely to develop brightness and saturation drift in the band. That matches the sampler literature: ancestral and SDE samplers never fully settle because they keep adding noise, while converging samplers walk a smooth deterministic path - exactly the path that shows a drift as a visible ramp.

Wiring and install

The node has no optional inputs and one output: overlap_params, a MMH3_OVERLAP_PARAMS link that goes to MMH3 Temporal Extend Video's overlap_params input. Use it or the Simple node, not both - they share the socket. Leave it unconnected and the Extend node falls back to its defaults (or the tile editor's legacy globals).

Install is the usual pack story - 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, no extra model files for this node specifically.

Practical order of attack

Start on MMH3 Temporal Overlap Simple and only come here when a specific failure needs a specific fix. Mosaic patterning at the join is fade_impl - move to qsample_init and expect to tune init_content_weight. A band that drifts bright or washed is a solver property before it's a knob, so change the sampler first.