MiniMax H3 Advanced Policy Override

MiniMax H3 Advanced Policy Override is the node you reach for when the default drift_av handoff between scenes isn't the one you want, and you don't want to rebuild the whole policy to change it. It's a named recipe you layer onto your existing Chain Policy - one input, one dropdown, one output. That's the whole node, and it's a genuinely useful way to experiment with scene transitions without touching anything else in the graph.

The reason it exists is that the 0.5 policy system packs a lot of decisions into one object: how motion carries across the scene boundary, whether generated audio continues, whether source audio is locked, and so on. The Advanced Policy doesn't try to reproduce all of that. It takes your chain_policy - from MiniMax H3 Chain Policy, another Advanced Policy, or the Legacy 0.4 adapter - and replaces only the incoming transition recipe, preserving the complete audio intent of whatever you fed in. If you need a later override to win, you can chain another Advanced Policy after it.

What the dropdown actually does

incoming_transition has ten choices, and they're the same vocabulary as the Plan's continuation_mode, so if you already know the difference between guide and masked_av, this feels familiar:

• cut - no continuity, a hard scene break.
• guide - the previous AV is supplied as fixed guide rows while the overlap regenerates.
• tone_guide, latent_guide, detail_guide - the experimental variants, progressively fancier handoffs.
• detail_av, drift_av, color_drift_av, hard_av, soft_av - the AV-mask family, where the predecessor's video tail is actually carried into the target latent and protected from denoising.

drift_av is the default and the one the maintained workflows use. color_drift_av is the new one in 0.5: it adds a tapered scene-one VAE delta to the copied video prefix for color stability, and - read the tooltip carefully - it does not mix with legacy 0.4 audio settings. The recipe also supplies its matching generated-audio context length, so you don't have to set audio_context_length by hand.

The status output is a plain STRING that reports the resolved transition and what audio intent was preserved. It's for eyeballing, not wiring.

Where you'd actually use it

The honest advice: start with the default drift_av and don't touch this node. The maintained workflows bake it in. Where you'll want it is when a specific scene boundary is misbehaving - say scene 2 keeps smearing into scene 3 - and you want to try soft_av or feathered_av on that chain without re-authoring the Plan. Because the audio intent survives, you can A/B the visual transition alone. That's the value: it isolates one knob.

One trap: because this is a layer, what you connect in determines what wins. Chain it after another Advanced Policy and the later one overrides the transition. Feed it the Legacy 0.4 adapter and it'll happily apply the advanced recipe on top - the tooltip explicitly lists that as a supported input, but the audio-interaction caveats get sharper there.

Installing it

Part of ComfyUI-MiniMaxH3-Contex-Loop. Install via ComfyUI Manager (search "MiniMax H3 Contex Loop") or:

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

Restart, and you'll find it under conditioning/minimax/contex_loop/policies. No extra dependencies beyond the pack itself - though, as ever, you need the H3 model weights and a ComfyUI build with native Add Guide support (PR #15439) for any of it to run.

Common issues

If your transition choice seems to have no effect, check you're actually on a 0.5 Chain Policy - this node's output has to feed Plan's chain_policy input (or chain onward). And if you mix color_drift_av into a workflow that still has a legacy 0.4 audio mode set, expect it to refuse; that combination is explicitly unsupported. When a scene boundary keeps breaking, don't fight it with prompts - try layering soft_av or feathered_av here first.