H3 Infinite - Output / Stitch v1.2

If you're chaining MiniMax H3 clips to make a long video, you've met the core problem this whole pack exists to solve: H3 only generates so many seconds per run, so every clip after the first starts by re-generating a "head" that overlaps the previous clip's tail. Stitch two clips together without trimming that overlap and you get a hard jump - 39 frames of picture that exist twice in a row, plus whatever frozen landing junk FL2VA left behind. H3 Infinite - Output / Stitch v1.2 is the node that does the accounting for you. It trims each rendered clip to exactly the frames the next clip actually reused, so your seams land on the same boundary the continuation latent used.

It's the middle of the "prepare each segment" stage: the Start and Continue nodes make the clips, this node prepares the rendered output, and then a Seamless Join or the Saved Chain stitcher assembles them. On its own it's just a smart trimmer - but it's the smart part of the whole chain.

How it works

The continuation machinery knows precisely where the reuse happened, and it records that in the H3_CONTINUOUS_HANDOVER metadata produced by the suite's Auto Handover analyzer. This node reads that metadata and cuts two things:

• the reused head context (the duplicated frames at the start of clip 2+, the overlap you must remove), and
• the freeze-safe tail (the last few frames where FL2VA locked up or drifted into a bright static landing - the part you never want to show).

The exact head_trim_frames and tail_trim_frames it computes come back as outputs, which is handy if you want to double-check its work or feed them elsewhere.

The inputs that matter

• images - the decoded FULL rendered frames for the current clip. Always wire the full decode, not an already-trimmed version.
• output_mode - the three-way switch that's the whole point:
  • Full keeps everything. True bypass, zero trimming.
  • Stitch Ready removes the reused head and the freeze tail. Use this for every intermediate clip.
  • Final Clip removes only the reused head, keeping the complete ending so the last segment reaches its full Last Frame landing.
• head_context_frames - the one people forget. Clip 1: leave it unconnected (0). Clip 2+: connect actual_head_context_frames from the Continue node, otherwise the duplicated head stays in your render.
• handover - the analyzer metadata. Required for Stitch Ready; without it the node throws rather than guess where the tail ends.
• audio - optional; the node trims audio to match the video trim.

Installing it

This is one node in Herrgotts-H3-Infinite-Continuation-Suite. Easiest path: ComfyUI Manager, search Herrgotts-H3-Infinite-Continuation-Suite, install, restart. Manual way:

cd ComfyUI/custom_nodes
git clone https://github.com/HerrgottMargott/Herrgotts-H3-Infinite-Continuation-Suite.git

Then restart ComfyUI and reload the browser UI. The pack ships no model weights - you still need the usual MiniMax H3 files (the pruned FL2VA checkpoint, Qwen3VL refs, video and audio VAEs) from Comfy-Org's repo. There are no pack-specific pip dependencies; the heavy lifting is ComfyUI's own stack. Note the license on the base model: MiniMax H3's community license excludes the US, EU, UK and South Korea, so check that before you start.

Common issues

• "stitch_ready requires H3_CONTINUOUS_HANDOVER metadata" - you forgot to connect handover, or you're feeding it a handover from the wrong clip. The tail cut is anchored to the analyzer's boundary, so it won't guess.
• Seams still jump on clip 2+ - almost always the missing head_context_frames wire. The head trim is what removes the double-exposed overlap.
• Full mode isn't trimming anything and you expected it to - that's the point. Full is a bypass; pick Stitch Ready or Final Clip.

It's an unglamorous utility node, but it's the one that decides whether your 11-clip masterpiece looks like one continuous shot or three re-exported clips taped together.