MMH3 Temporal Overlap Simple

MMH3 Temporal Overlap Params exposes thirteen knobs for how one H3 segment hands off to the next. Thirteen knobs is a lot when you just want to extend a clip. This node is the answer: one preset dropdown and one fade_frames field, feeding the exact same overlap_params socket on MMH3 Temporal Extend Video.

That's the whole node. It samples nothing, loads nothing, and has no optional inputs. Pick a preset, wire the output, move on.

The four presets, and what they're really doing

Read these as a trade-off table, not a difficulty ladder:

high continuity - smooth, freeze+fade gradient mask at full mask strength with a hard seam anchor. This is the stock configuration expanded: maximum preservation of the carried tail, anchor_strength 0.999, the earlier content winning the overlap band with a linear blend. The catch is in the description itself: quality softens over many rounds, because a chain of frozen-then-faded joins is a chain of content the model never gets to re-render.

low deterioration - near-no deterioration with still-smooth transitions, at the cost of content consistency, which "tends to jump". Under the hood it's a weak mask (strength 0.2) with the seam anchor switched off. So less is held rigidly in place - the frames can breathe, and don't bleach - but there's nothing pinning the composition across the join.

soft restart - barely any deterioration, the carried tail re-samples, and the new segment overwrites the band. Contiguity is rough, with a tiny seam jump. It's free_resample tail mode, a softened anchor (0.6), the new segment winning the overlap, and no blending at all: overwrite.

H3 video reference - a different mechanism entirely. Instead of the fade mask carrying continuity, it uses H3's native reference-video feature: the prev_tail seam reference spanning the whole carried tail (0 = all of it). The preset also drops the previous segment's keyframes and overwrites the band. Two honest warnings come with it: it needs prompting to cooperate - the model has to be told what to do with the reference, it won't just follow it - and it samples much slower.

If you're starting from scratch: high continuity for a two- or three-segment chain, because it looks best while deterioration hasn't had time to build. For a long chain, low deterioration or soft restart, and accept that you're trading seam smoothness for the clip not getting progressively softer.

What fade_frames does here

It's the length in pixel frames of the fade band between the carried tail and the new content, and it has to fit inside that segment's overlap frames from the tile editor. 0 means no fade - the whole tail stays frozen and the new content hard-cuts in, which is a legitimate choice when the reference is doing the work instead. Non-zero values snap up to multiples of 17, matching H3's keyframe grid.

Two of the four presets ignore this field: soft restart and H3 video reference both re-sample the whole tail, so there's no frozen band to fade across. The author's note is that H3 video reference is designed for 0 anyway, since the seam reference already carries the continuity.

A detail that builds trust

Presets here are complete parameter sets, not partial overrides. Each one is expanded against a full base dict, so the preset alone decides the continuation behaviour - a leftover value sitting in the tile editor's legacy global extend-params can't leak in and quietly change what "high continuity" does. If you've ever debugged a preset that wasn't actually doing what it said, you know why that matters.

It also means the two overlap nodes are genuinely interchangeable at the socket: same output type, and whichever one you connect wins. Use one, not both.

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 Python dependencies to install. You need a ComfyUI with MiniMax H3 support built in, plus the H3 model, CLIP and VAEs.

When it doesn't save you

Presets can't fix a sampler problem. If the fade band is drifting bright or desaturated across a chain, the cause is usually the sampler rather than the preset: the pack recommends stochastic SDE-type samplers - sa_solver, er_sde, dpmpp_2m_sde, dpmpp_3m_sde - and warns that deterministic ones (euler, res_multistep, uni_pc, dpmpp_2m) tend to develop brightness and saturation drift in the band. No preset here changes which sampler you selected, so if a chain degrades in that specific way, that's the first thing to change.

And if none of the four clicks, the full MMH3 Temporal Overlap Params node carries every value these presets expand to, including the qsample_init fade implementation that fixes the mosaic artifacts the default mask band can produce on H3. The presets are the shortcut, not the ceiling.