MiniMax H3 Contex Loop Context

MiniMax H3 Contex Loop Context is the heart of the continuity system. Everything the loop promises - "your character still looks like the same person in scene four," "the room tone carries," "the camera doesn't reset" - comes down to what this node injects into the conditioning and latent before each scene samples. If you're going to understand one node in this pack deeply, make it this one.

It sits between the stock MiniMax H3 Ref2VA/I2V conditioning node and your sampler. In scene one it's a pass-through: no predecessor exists, so your conditioning goes through untouched. From scene two on, it applies whatever context the Plan and Chain Policy called for - the saved video tail as guide rows, or an AV-mask mode where the previous tail is VAE-encoded into the target latent and protected from denoising.

The inputs

• state - from Current Shot. This is the scene bookkeeping: which scene is rendering, what the Plan resolved.
• conditioning - from the stock H3 node. Scene 1 passes through; later scenes get the saved continuation context attached.
• vae - the H3 video VAE, used to encode saved context frames for continuation scenes.
• latent - the current scene's empty AV latent from the stock conditioning node. In guide mode it passes through; in AV-mask modes it comes back as a masked preserved-prefix copy.
• audio_vae (optional) - the H3 audio VAE, required when scene 1 continues from imported audio or when Chain Policy locks source-audio windows into the target latent. You can leave it disconnected only if neither behavior is active.
• model and drift_sigmas (optional) - only for Drift-Control AV and Color-Stable Drift AV. For a single-model sampler, connect one H3 MODEL here. For a sigma split that switches models mid-schedule, leave MODEL disconnected, connect the full drift_sigmas schedule, and put a Drift-Control Model Patch on each raw model branch.

The outputs that matter

• conditioning - sampler-ready. Connect to the H3 guider.
• trim_frames - the repeated leading frames to strip after decoding. This is the number that keeps your delivered scenes the right length; wire it to MiniMax H3 Contex Loop Trim.
• latent - the sampler-ready target latent. In AV-mask modes this is where the preserved video prefix lives, and with Lock source audio it carries the complete scene-local audio stream even in Guide/Cut modes. Wire this output to the sampler, not the original latent.
• is_continuation - a BOOLEAN that's true whenever preceding video or generated audio is carried, including audio-only guide continuation. Handy for a conditional in the graph, but honestly most people use it for debugging.
• model - the H3 MODEL, patched only while a Drift-Control AV mode is active; otherwise your connected model passes through unchanged.

The honest read

For a normal first run you don't touch much: connect state, conditioning, the video VAE, and the current latent, and let the Chain Policy drive the modes. Where people get burned is mixing up the two VAEs - the video VAE goes in vae, the audio VAE in audio_vae, and swapping them produces the kind of nonsense decode that's genuinely hard to diagnose. Also note the latent output: in 0.5 you must wire the sampler to the returned latent so the Plan can switch modes safely, not the original one.

Installing it

Ships in 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. You need the H3 model weights and a current ComfyUI build with native Add Guide support (PR #15439). ffmpeg is preferred for the review/assembly steps around it; PyAV works as fallback.

Common issues

Besides the VAE swap above, the classic failure is an AV-mask mode with a context_length that doesn't hit the AV clock. The Plan tooltip is blunt: use 22 for guide mode and 39 for masked/tapered/feathered AV so the clocks meet exactly. If a scene comes out with a frozen-looking head or the wrong duration, check context_length, trim_frames wiring, and whether Chain Context's latent output actually reaches the sampler.