MiniMax H3 SPEED Stage Source / 分阶段条件源 (Advanced)

Here's the subtle problem SPEED creates: if you sample at half resolution first and full resolution later, then every stage needs its own conditioning, re-encoded to that stage's canvas. A first-frame image encoded once at 1056×608 is the wrong shape at 528×304. MiniMaxH3SPEEDSourceT8Advanced is the node that solves it by holding onto the raw text and media inputs - the untouched prompt, first_frame, last_frame, reference images, reference videos, and reference audios - and packaging them into an H3_T8_SPEED_SOURCE that the SPEED sampler can resize and re-encode per stage. It doesn't load a second H3 model, and it doesn't touch the stable Conditioning node you already use for normal runs.

What it does

Think of it as a "conditioning source vault." You wire in the things you'd normally hand to a conditioning node - clip (native H3 Qwen3-VL), video_vae, audio_vae, the prompt, and any media references - plus the task contract: length (124 frames default, snapped to the 17n+5 grid at 24fps), task_type (auto lets the pack decide between T2VA / I2VA / FL2VA / L2VA / Ref2VA / Hybrid), and audio_mode (native, lock_source, remix_source, or reference_only).

The reference inputs are autogrowing: ref_images, ref_videos, ref_video_audios, and ref_audios each expand to hold up to a handful of items, and the pack's strict-prompt-tag system (strict_prompt_tags, default on) makes sure your <Picture N> / <Video N> / <Audio N> labels in the prompt actually match what's connected. prompt_primary_audio_ordinal (default 1) marks which <Audio N> is the primary voice reference.

The fingerprint inputs (checkpoint_fingerprint, vae_fingerprint) matter only for the delta-optimal SPEED path - if you're on manual sigmas (which you should be, starting out), leave them at unrecorded. The tooltips say it plainly: "Manual-sigma plans do not require it."

Why not just reuse the stable Conditioning node?

The stable node conditions once, for one canvas. SPEED stages at multiple resolutions, so each needs its own re-encode - of the first frame, the last frame, the references, and the keyframes. The Source node keeps those originals around precisely so the sampler isn't stuck re-encoding from already-downscaled latents. The description's phrasing is the whole design: "every SPEED stage can resize and re-encode keyframes/references for its own canvas."

Installing and pairing

Same pack install: ComfyUI Manager → "MiniMax H3 Audio T8", or git clone https://github.com/T8mars/comfyui-minimax-h3-audio-T8 into custom_nodes, restart. No pip deps; the H3 weights and your reference media are yours. In the graph it sits between Plan and Sampler: speed_plan → Sampler, speed_source → Sampler. This is experimental, so expect the strict-prompt-tag validation to be your most common error source - when it complains, it's usually a real mismatch between your <Audio N> labels and the actual inputs, and the README's "prompt media tag validation failed" note is your troubleshooting map.