MiniMax H3 Director Planner Conditioning

MiniMax H3 Director Planner Conditioning is the Director Conditioning node with one extra output, and that one output tells you something the plain version keeps to itself: which mode H3 actually resolved your inputs into. Where its sibling just produces conditioning and walks away, this node adds a third task_mode output - a short string like r2v - Reference to Video (MiniMax H3) (~2 ref image(s)) - so a planning layer downstream can see what the graph is doing.

Everything else about the inputs is identical to MiniMax H3 Director Conditioning: clip, vae, prompt, width, height, length, optional audio_vae, first_frame/last_frame keyframes, up to nine reference_image_N slots, and ref_image_size (match/max). No task-type dropdown here either - the node watches what's connected and infers. Reference images present → it runs H3's MiniMaxH3ReferenceToVideo path (and demands the audio_vae); otherwise it runs MiniMaxH3ImageToVideo. The task_mode output is just that inference, reported honestly, including a reference count when one applies.

Why the extra output exists

H3 is a chameleon: the same prompt, width, height, and length describe a text-to-video clip if nothing else is plugged in, an i2v/fl2v clip if keyframes are, and an r2v clip if reference images are. When you're scripting a graph from another node - a planner, an LLM prompt-routing setup, anything that builds clips on the fly - you need to know which family the conditioning landed in so you can pick the matching UNET and decode path downstream. That's this node's reason to exist: it surfaces the mode decision as data you can branch on, instead of leaving it implicit.

It's also a decent sanity check during setup. Wire one up, queue once, and read task_mode to confirm your reference images actually pushed the graph into reference territory rather than silently producing a plain t2v clip.

Outputs

positive (CONDITIONING) and latent (LATENT) feed a KSampler on the H3 UNET with a MiniMaxH3SigmaShift, exactly like the sibling node; task_mode (STRING) is the human-readable mode report for planning. The decode-side caveat carries over too - this node does conditioning, not the AV-latent separation that turns your sampled result into video plus stereo audio, so budget for that downstream if you're building a full graph.

If you're doing real multi-segment directing, though, step back and ask whether you need this at all. The Director node does conditioning, sampling, and decode per segment and hands you finished images and audio - it just doesn't expose the conditioning as reusable pieces. Planner Conditioning is the tool for people who need those pieces in the open. ComfyUI 0.30.0+ is required (it calls ComfyUI's official H3 nodes), and the H3 Community License territory restriction - US, EU, UK, South Korea - applies to whatever you generate with it.