MiniMax H3 Guide Model

You didn't add this node. The pack did.

If you opened a Continuity workflow and found "MiniMax H3 Guide Model" in the middle of the graph, don't try to wire it by hand - the pack marks it internal and writes it into the graph the moment you switch on the guide LoRA pill on H3's sampler row. Its whole job is one step of the guide pass: put a stack on the checkpoint.

The guide pass re-generates a finished clip with the clip itself pinned as an aligned guide, under a file trained to map one video to another. Those files are LoRAs - Alissonerdx's minimax_h3_lms sharpener and minimax_h3_style_transfer, rank-64 Ref2VA adapters trained on ostris's ai-toolkit fork - and the loader that puts them on matters more than you'd guess.

Why it's a node, and where it sits

Because of what comes after the LoRAs. The stack goes on first, then the sigma shift, the accelerator and the preview decoder - the order every other sampler in this family runs behind - and a block-cache accelerator refuses to sit downstream of a block replacement. Hence a node rather than a line hidden inside the pass.

Why it forces ComfyUI's own loader

This is the reason it exists rather than reusing the pack's own LoRA machinery. Continuity normally patches H3 adapters with a vendored stack that keeps the quantized base intact and runs the delta as an exact bf16 branch. The guide pass overrides that with core's loader, the one that dequantizes a touched layer, adds the delta and requantizes with fresh rounding.

That looks like a downgrade. It isn't. Rebuilding the published style-transfer workflow out of core nodes, the author measured: with core's loader the file moves the picture; with the exact vendored stack on the same graph the result comes back near-photoreal, meaning the style file barely engaged. Why requantizing transfers more isn't understood. What is understood is that these files were trained, judged and published against core's loader, so a pass whose promise is "the published result" has to wear them the way the publication does - which means your settings.lora_loader choice for the rest of the pack never reaches this node. Deliberately.

The inputs that matter

model takes the checkpoint the pass asked for - Ref2VA by default, since that's what the guide files were trained against (their cards say FL2VA works too, less tested). checkpoint names which of H3's two that is, and does double duty: it's what the pack requests at queue time, and it settles the claims - an entry claiming only FL2VA is skipped when Ref2VA is the target, and vice versa, since the two are different weights and a LoRA trained on one does nothing on the other.

loras is the stack as JSON, per the tooltip: the piece's turbo distill where it has one, then the guide file. In practice:

[{"name": "minimax_h3_ref2v_turbo_4step_v0.1_comfyui_bf16.safetensors", "strength": 1.0},
 {"name": "minimax_h3_lms_v1.0_r64.safetensors", "strength": 1.0}]

Entries with strength of 0 or enabled: false are dropped, and the guide file's strength is clamped to the 0–2 range the published card uses. The single MODEL output goes into whatever patches sample on it, then into the guide pass.

Install

cd ComfyUI/custom_nodes
git clone https://github.com/roadmaus/ComfyUI-Continuity

Restart ComfyUI - nothing to pip install, and Manager users can just search the pack title. The node needs no download, but the files it loads do:

• the guide LoRAs in models/loras: minimax_h3_lms_v1.0_r64.safetensors and minimax_h3_style_transfer_v1.0_r64.safetensors from Alissonerdx/Minimax-H3-ComfyUI
• the distill the pass is meant to wear, a minimax_h3_ref2v_turbo_4step file, same folder
• the H3 checkpoints, text encoder and VAEs the render needs (see the pack's docs/models.md)

Where people get burned

• "The guide-LoRA pass was given no file to wear." You queued with the pill on and an empty stack.
• "LoRA not found in models/loras: <name>". The JSON names a file that isn't there - usually one hand-typed from memory.
• Swap the distill and the pass halves. Without the published ref2v…turbo file the piece's own turbo stands in - measured as under-driven: a style by half, a sharpen barely.
• Two copies of this pack and no node shows up at all. The pack was renamed from ComfyUI-MiniMax-Creator and the node ids stayed the same, so an old folder beside the new one registers those ids twice and ComfyUI shows neither. Delete one, restart.
• "Render refused, naming a field and a folder" isn't this node - a weight file is missing somewhere in the render. Put the named file in the named folder.

And expect roughly double the wall clock - the pass is a second full generation per shot. Nothing extra gets loaded, but if you're OOM-ing on long renders with recent ComfyUI, the pack's FAQ fix is starting with --disable-dynamic-vram.