Nodes/MiniMax H3 RefDelta Solver/[Legacy] MiniMax H3 RefDelta Reference Guider
ComfyUI Node

[Legacy] MiniMax H3 RefDelta Reference Guider

The old two-model diagnostic that VRAM killed

By xmarre·Created 4 days ago·Updated 2 days ago· 1
[Legacy] MiniMax H3 RefDelta Reference Guider
  • model
  • reference_model
  • positive
  • negative
  • guider
cfg1.0

What this is

A deprecated diagnostic guider that runs two full H3 models at once - the fused RefDelta checkpoint and a genuine reference model - on the same latent and conditioning, and compares their predictions as sampling goes. It was the original way to check how close the fused model's x0/velocity is to a real model on the same state. The display name says it: [Legacy].

If that setup sounds expensive, you're right, and that's exactly why it's legacy.

How it worked

It's a CFGGuider subclass with a reference-model sidecar. You feed it your fused model, a distinct reference_model (it refuses to run if model is reference_model - it raises a ValueError), positive and negative CONDITIONING, and a cfg (default 1.0, the usual guidance-distilled territory for H3). Each sampling step, the guider computes the fused prediction and the reference prediction on the same state and records the difference - cosine, magnitude, projection, residual and orthogonal components, split per video/audio stream.

Output is a GUIDER, which plugs into the guider socket of a SamplerCustomAdvanced-style custom sampling setup. The reference model is handled with its own model patcher, offloaded and cleaned up after the run.

Why you shouldn't use it

Two reasons, both in the node's own description:

  • It may require both full H3 models resident in VRAM simultaneously. H3 is a 33B model; two of them is a memory problem most people don't have. The fused checkpoint alone is a big download.
  • The comparison replay path replaced it. The README's recommendation is blunt: "prefer diagnostic capture + Comparison Replay." The newer MiniMax H3 RefDelta Comparison Replay sampler achieves the same same-state comparison while loading one full model at a time, across three passes, and it's the maintained, tested path.

This node is kept registered purely so saved workflows that used it still load.

Install

cd ComfyUI/custom_nodes
git clone https://github.com/xmarre/ComfyUI-MiniMax-H3-RefDelta-Solver.git

Restart ComfyUI; no pip dependencies beyond current ComfyUI. If a workflow you downloaded still uses this node and you just want to generate, swap it for the normal guider setup - this one is for diagnostics, and even then it's the old way.

Common issues

  • ValueError: requires a distinct reference MODEL - you wired the same model into both sockets. They have to be different checkpoints.
  • Out of memory - you loaded two 33B H3 models at once. This is the expected failure mode, and the reason the node is legacy.
  • It's not a quality tool. Like the rest of the pack's reference machinery, this compares states to understand the imported delta approximation; it was never a scheduler or quality oracle.
Categorysampling/custom_sampling/guiders

Inputs (5)

NameTypeDefaultDescription
modelMODEL
reference_modelMODEL
positiveCONDITIONING
negativeCONDITIONING
cfgFLOAT1.00–100

Outputs (1)

NameTypeDescription
guiderGUIDER