Nodes/H3 Continuum/[Legacy] H3 Continuum Join
ComfyUI Node

[Legacy] H3 Continuum Join

The V1 building block that chains one H3 clip to the next

By ukr8b3g-cmyk·Created 29 days ago·Updated a day ago· 73
[Legacy] H3 Continuum Join
  • model
  • conditioning
  • latent
  • previous_state
  • model
  • conditioning
  • latent
  • plan
  • report
continuityBalanced — 22 frames
extend_seconds5.0
audio_continuityfalse
first_frame_policyUse Continuum context; keep visual identity tokens (Recommended)
preserve_last_frametrue
strict_compatibilitytrue
debugfalse

H3 Continuum Join is the workhorse of the pack's V1 generation - the manual, block-by-block way to build a long MiniMax H3 video. The modern samplers chain chunks for you automatically; this node is the Lego brick those samplers replaced. You place it in front of a normal KSampler, and it prepares a continuation clip: it takes the previous run's audio/video latent state, positions it on the new clip's timeline as one context reference, and gives you a model, conditioning, and latent that you feed into your regular sampler to generate the next segment. Then Finish trims that segment and captures the state for the next Join, and Assemble glues the decoded segments together.

Because it slots into a normal sampling graph, it composes with whatever you already have in the chain - the description is explicit that it uses per-MODEL wrappers only and works alongside external SageAttention, Sol-Attn, and Spectrum.

Inputs that matter

  • model, conditioning, latent - the normal KSampler wiring, transformed by the node into continuation-ready versions.
  • previous_state (optional) - the H3_CONTINUUM_STATE from the previous Finish. Without it, the node has no context and can't continue.
  • continuity - Balanced - 22 frames (reproduces the V1 default), Fast - 5 frames, or Strong - 39 frames (Experimental). This is how much prior video context the next clip gets.
  • extend_seconds (5.0) - the new duration requested. The tooltip is the key detail: H3 snaps to the nearest 17k+5 frame count, so the actual output won't always be exactly what you typed.
  • audio_continuity (true) - carry the previous native audio latent and place it on the new clip's timeline.
  • first_frame_policy - default keeps the Continuum context while preserving visual identity tokens; alternatives error on an existing keyframe or keep the first-frame keyframe too (advanced).
  • preserve_last_frame (true) - moves an existing FL2VA last-frame anchor to the end of the continuation clip.
  • strict_compatibility (true) - proactive native-H3 API checks; marked payload failures always stop.

Outputs: model, conditioning, latent (into your KSampler), plus plan (an H3_CONTINUUM_PLAN that Finish consumes) and report.

Install

cd ComfyUI/custom_nodes
git clone https://github.com/ukr8b3g-cmyk/ComfyUI-H3-Continuum.git

or via ComfyUI Manager ("H3 Continuum"), then restart. No extra pip packages; needs a MiniMax H3 ComfyUI environment (weights not bundled) and ComfyUI v0.32.0+.

Honest verdict: this is the fiddly old way, and the current V3.3 sampler exists precisely so you don't have to string Join → KSampler → Finish → Assemble by hand across a dozen clips. Keep this node in your vocabulary for reading old workflows - the manual chaining loop is the clearest explanation of what "continuation" means in this pack - but reach for the integrated sampler for actual work. If you do build with it, the loop is: Join (with the previous Finish's state) → sample → Finish → next Join. Forget the previous_state wire and you're just generating unconnected clips.

CategoryMiniMax H3/Continuum/Legacy

Inputs (11)

NameTypeDefaultDescription
modelMODEL
conditioningCONDITIONING
latentLATENT
continuityCOMBOBalanced — 22 frames3 options: Balanced — 22 frames, Fast — 5 frames, Strong — 39 frames (Experimental)
extend_secondsFLOAT5.00.5–15Requested NEW duration. H3 snaps to the nearest 17k+5 frame count.
audio_continuityBOOLEANfalseCarry the previous native audio latent and place it on the new clip timeline.
first_frame_policyCOMBOUse Continuum context; keep visual identity tokens (Recommended)Recommended: replace only the old spatial keyframe. If the prompt was encoded with an input image, its Qwen visual tokens remain as an identity cue.
preserve_last_frameBOOLEANtrueMove an existing FL2VA last-frame anchor to the end of the continuation clip.
strict_compatibilityBOOLEANtrueRun proactive native-H3 API checks. Marked Continuum payload failures always stop; they are never silently bypassed.
debugBOOLEANfalse
previous_stateoptH3_CONTINUUM_STATE

Outputs (5)

NameTypeDescription
modelMODEL
conditioningCONDITIONING
latentLATENT
planH3_CONTINUUM_PLAN
reportSTRING