MiniMax H3 Conditioning

If the KSampler is the engine of an H3 workflow, this node is the fuel line. It's the one place where the whole prompt gets turned into something the model can actually use: your text goes through the Qwen3-VL-32B encoder, your reference images or videos get encoded through the video VAE, and a blank joint audio-video latent gets created at the resolution you ask for. Three outputs - positive, negative, latent - and they feed exactly one consumer: MiniMax H3 KSampler.

How it works

Everything comes together in make(). The text_encoder input runs the prompt through Qwen3-VL; the av_encoder input encodes any reference media you've attached; an FL (first/last-frame) constraint gets folded in if present; and the node builds the MINIMAX_H3_LATENT - a zeroed tensor sized [1, 24, latent_t, H/16, W/16] for video plus [1, 32, 2, audio_t] for stereo audio at 40 fps latent rate. The latent's frame count comes from the prompt payload's total_duration (via frame_count), so the prompt node is what decides clip length, not this node.

One detail worth knowing: the canvas isn't raw pixels. The node uses a 768-short-edge canvas capped at the 768×1344 area - it picks the nearest 32-multiple width/height that fits that area for your ratio. So you can type 1344×768 and get exactly that, but a wild ratio gets normalized rather than honored literally.

The inputs that matter

• text_encoder (required) - from MiniMax H3 Encoder Loader. Without it, nothing happens.
• width / height (required) - output latent resolution, in multiples of 32. Defaults 1344×768. These set the render resolution, full stop.
• prompt (optional) - a MINIMAX_H3_PROMPT from SimplePrompt, Storyboard, or a Refiner. If a text prompt is also wired elsewhere, the payload's text wins.
• package (optional) - PACKAGE_DATA from the PackageData node, for reference images/videos/audio. Mutually exclusive with fl_constraint - the README is blunt: you can't use both in the same Conditioning node.
• fl_constraint (optional) - from the FL Constraint node, for first/last-frame modes.
• av_encoder (optional) - from the VAE Loader. Needed for reference modes; harmless to wire always.
• ref_max (optional) - "Reference longest-edge pixel limit before encoding", default 1280. Caps how big your reference media gets before the VAE encodes it; lower it if reference encoding eats your VRAM.

The outputs

positive and negative - both MINIMAX_H3_COND - plus latent. All three go into the KSampler. Note the negative side: it only means something when the KSampler's cfg is above 1.0. At cfg=1.0 (the default) negative guidance is off and the negative conditioning is effectively decorative.

Installing it

Same pack routine - ComfyUI Manager search "MiniMax H3", or:

cd ComfyUI/custom_nodes
git clone https://github.com/xiaolibai-sys/ComfyUI-MiniMaxH3
pip install -r requirements.txt

then restart. No model files are needed for this node itself - it consumes whatever loaders feed it.

Where people get burned

• Latent resolution ignored. If you set width/height and nothing changes, check whether the prompt payload forced something - or remember the canvas cap: an area over 768×1344 gets scaled down. That's H3's model-side limit, not a bug.
• "Can't use FL and PackageData together." It's a real guardrail, not a suggestion. Pick reference media or first/last frames for a shot.
• Negative conditioning has no effect. Your cfg is 1.0. Raise it above 1.0 (and the KSampler will then require the negative socket to be filled).