Nodes/ComfyUI-MiniMaxH3-Contex-Loop/MiniMax H3 Chain De-Rope Continuity
ComfyUI Node

MiniMax H3 Chain De-Rope Continuity

Splice the previous HQ tail back in before pass 2 re-times the scene

By ethanfel·Created 17 days ago·Updated about 19 hours ago· 315
MiniMax H3 Chain De-Rope Continuity
  • state
  • video_latent
  • video_latent
  • status

MiniMax H3 Chain De-Rope Continuity is the last piece of the de-rope puzzle, and it handles the one case the other two guards can't: a Drift-Control scene whose continuation prefix needs to be the previous scene's high-quality tail, not the current target-resolution copy.

Here's the setup. In the deferred upscale loop, each scene is upscaled separately. A Drift-Control scene 2+ carries a continuation prefix - the last frames of the previous scene - and that prefix exists in two forms: the saved HQ (high-quality, already-upscaled) tail from the previous scene's completed pass, and the target-resolution copy that got fed through the current scene's upscaler. For pass 2 to continue from genuinely high-quality continuity, the target-resolution prefix should be replaced with the previous scene's saved HQ latent tail. If you skip that, your scene 2 continuity is literally lower-resolution than the pixels around it - a visible quality cliff right at the scene boundary.

What it does

It takes the current Upscale state (which includes the prior HQ context, restored for a resumed child run) and the video_latent - the target-resolution 24-channel clean video latent after Time Smear, VAE encode and the spatial latent upscaler. For Drift-Control scenes it splices the previous scene's saved HQ tail into the protected prefix of that latent. Every other scene passes through unchanged. The video_latent output then goes to MAINodes H3 V2V Init.

The status output tells you which branch happened: whether prior HQ context was spliced, source context was protected, or the scene has no latent continuation prefix at all. That's a genuinely useful diagnostic - it tells you whether the node considered the scene a continuation.

The honest read

This node is thin in interface (two inputs, no settings) and deep in consequence. If you run the bundled "Deferred Upscale + De-Rope - H3 LBH 3D" workflow with Drift-Control scenes and leave this node out, you'll get boundary artifacts you'll blame on the upscaler. It's also the node that makes resuming an upscale child run coherent: the state carries the prior HQ context so a resumed run restores continuity instead of starting fresh.

Where it fits

The full chain, from the pack's docs:

source x0 → H3 Jerk Oracle → Chain De-Rope Guard → H3 Time Smear
          → H3 video VAE encode → LBH 3D → Chain De-Rope Continuity
          → H3 V2V Init + Inject Schedule → sampler

Guard protects timing, Freeze Mask protects texture, Continuity upgrades the prefix to HQ. Three nodes, three different jobs, one boundary that survives.

Installing it

Part of ComfyUI-MiniMaxH3-Contex-Loop (ComfyUI Manager: "MiniMax H3 Contex Loop"), or:

cd ComfyUI/custom_nodes && git clone https://github.com/ethanfel/ComfyUI-MiniMaxH3-Contex-Loop.git

Restart ComfyUI. It wraps MAINodes H3 V2V Init, so install ComfyUI-MAINodes (matlowai) too. Standing requirements: H3 weights, current ComfyUI with native Add Guide (PR #15439).

Common issues

If the boundary between a Drift-Control scene and its predecessor still shows a quality jump, check status - "source context was protected" means the node didn't find a saved HQ tail to splice, which usually points at the previous scene not being a completed Drift-Control pass or the checkpoint missing the saved denoised latent. And note the splice only applies to Drift-Control scenes; for other continuation modes the node intentionally passes through unchanged.

Categoryconditioning/minimax/contex_loop/upscale

Inputs (2)

NameTypeDefaultDescription
stateH3_CHAIN_UPSCALE_STATECurrent Upscale state, including the prior HQ context restored for a resumed child run.
video_latentLATENTTarget-resolution 24-channel clean video latent after Time Smear, VAE encode, and the spatial latent upscaler.

Outputs (2)

NameTypeDescription
video_latentLATENTTarget-resolution de-rope video init with the prior HQ tail spliced into a Drift-Control prefix when required.
statusSTRINGWhether prior HQ context was spliced, source context was protected, or the scene has no latent continuation prefix.