MiniMax H3 Director Conditioning

ComfyUI's official MiniMax H3 nodes bundle a lot of work into "conditioning": they take your text, encode it with the Qwen3-VL CLIP, stuff any keyframes or reference images into the picture, and produce the positive conditioning plus the audio-video latent that sampling actually consumes. That stage is normally invisible - the Director swallows it whole. MiniMax H3 Director Conditioning is that stage pulled out and exposed as a plain node, for the times you want to hand-build a single-clip graph instead of renting the Director's whole timeline.

Think of it as the Director pack's version of what the official workflow templates wire up between the model loaders and the KSampler. Same result, more control, fewer walls.

The inputs

Straightforward model plumbing up top: clip (a minimax-type Qwen3-VL CLIP), vae (the H3 video VAE), plus prompt, width, height, and length - that's frames at 24fps, default 124 ≈ 5s, and it's stepped to H3's frame grid. Then the optional layer:

• audio_vae - required the moment you go down the reference path. Missing it with reference images connected is the most common hard error here.
• first_frame / last_frame - keyframes for i2v/fl2v work.
• reference_image_0 through reference_image_8 - up to nine reference images, referenced in the prompt as <Picture 1>… <Picture 9>. Wire any of them in and the node switches to the reference-to-video machinery.
• ref_image_size - match or max, how reference images get sized at encode time.

What you won't find is a task-type dropdown. The node infers its path from what you plug in: any reference image connected and it runs H3's MiniMaxH3ReferenceToVideo conditioning (which is why audio_vae then becomes mandatory); otherwise it runs MiniMaxH3ImageToVideo. That auto-selection is convenient until it surprises you - connect a reference image you meant as a style hint and you've silently changed the task family.

Outputs and where they go

Two outputs: positive (CONDITIONING) and latent (LATENT). These are the exact hand-offs the Director feeds its sampler, so your graph continues the same way the official H3 pipeline does: into a KSampler running the H3 UNET with a MiniMaxH3SigmaShift applied, then decode. That decode side is the part this node doesn't cover - H3's latent splits into picture and stereo audio, and you need the official AV-latent separation to get sound out. Plan your downstream before you assume conditioning + a sampler is the whole job.

When to use it - and when not to

Reach for it when you want a single clip with bespoke plumbing: a custom sampler, an experimental scheduler, per-clip control you can't express inside the Director's one-stop interface. Skip it if you're doing multi-segment work - that's what the Director node is for, and this node adds graph overhead without giving you the timeline, segment continuity, or the images/audio outputs you'd have to rebuild by hand.

ComfyUI 0.30.0+ is non-negotiable (it calls ComfyUI's official H3 conditioning nodes directly). If you're in a region the H3 Community License excludes - US, EU, UK, South Korea - that applies here too, since this node is only as legal as the weights behind it.