MiniMax H3 All-in-one Conditioning - Star7

The fiddly part of MiniMax H3 in ComfyUI isn't the sampling - it's the front end. H3 is an omni-modal model that treats text, images, video and audio as one context, which is great until you're hand-assembling a conditioning stack where your first frame, two reference videos, three reference images and a drive-audio track all have to land in exactly the right placeholders in exactly the right order. MiniMaxH3MaterialPromptStar7 collapses that whole front end into one node, the way the good SD-pack conditioners did for text, and it adds a bonus: it hands its context forward to the face-repair node so you never build the stack twice.

It's part of star7code's minimax-h3-chunk-star7 pack - the same repo whose headline node does H3 activation chunking and selectable attention. This one is the "All-in-one Conditioning" helper that slots in front of your sampler.

What it does

Feed it the raw ingredients - your loaded model, clip, video_vae, and audio_vae (yes, H3 uses a separate audio VAE; this node expects both), plus a prompt and whatever visual/audio material you have. It then does all the bookkeeping a hand-built stack would make you do yourself:

• Normalizes prompt placeholders into H3's official tags - <Picture 1/2> for connected reference images, <Video 1/2> for reference videos, <Audio 1/2> for reference audio - and converts the dedicated driver-audio alias <Audio D> into the model's internal numeric id. You can write those tags naturally and the node reconciles them against what's actually wired.
• Snaps width/height to the nearest 32-pixel multiple and handles reference frames internally at canvas resolution.
• Aligns reference video to H3's 17n+5 frame grid at 24 fps, auto-trims anything past 15 seconds or beyond the generated length, and crops companion audio tracks to match.
• Routes the audio latent according to your audio_mode and encodes first/last frames, reference images and videos into the packed latent.

Outputs are model (still conditioned/patched), positive CONDITIONING, the av_latent your sampler consumes, mux_audio (the audio track to mux into the final video), a report string with what it decided, and - the star of the show - refine_context, a STAR7_H3_REFINE_CONTEXT you plug straight into MiniMax H3 One-click Face Repair - Star7 so that node reuses this same model and conditioning instead of rebuilding them.

The inputs that matter

• task_type - leave on 自动判断 / Auto and it infers text-to-video, first-frame, first-last-frame, reference-material, or hybrid from what you connected. The explicit modes (T2VA / I2VA / FL2VA / L2VA / Ref2VA / Hybrid) exist for when auto guesses wrong.
• width / height / length - defaults are 1344×768 and 243 frames (~10 s at 24 fps). Note the frame stepper is 17, H3's grid.
• audio_mode - 锁定原音 (lock source audio), 重混原音 (remix source), 仅作音频参考 (reference only), or 模型原生生成 (let the model generate audio natively). There's no hidden audio-enhancement model here; stability comes from sample-rate conversion and proper latent routing.
• audio_denoise_strength - how hard audio reference conditioning bites, default 0.35.
• reference_quality - match the generation canvas (cheaper) or keep full-resolution references (more VRAM).
• Optional inputs: first_frame, last_frame, ref_image_0..3, ref_video_0..2 (each an IMAGE batch - feed video as frames), ref_video_audio_0..2, ref_audio_0..1, plus drive_audio and final_audio.

Everything is optional beyond the four core model/VAE sockets and the prompt, so a pure text-to-video run is just model, clip, both VAEs and a prompt.

Installing it

Same pack as its sibling nodes - install once, get them all:

cd ComfyUI/custom_nodes
git clone https://github.com/star7code/minimax-h3-chunk-star7.git

Restart ComfyUI afterward, or find "MiniMax H3 Activation Chunk - Star7" in ComfyUI Manager. Runtime deps (scipy, scenedetect, ultralytics) install automatically. The H3 weights themselves are not in this pack - you load those separately, and this node just conditions what you loaded.

Gotchas worth knowing

• It expects a real H3 audio VAE, not a reused video VAE. Missing audio_vae connections are the usual first-run stumble.
• A reference video under ~5 frames won't run; under 2 seconds it only warns. Over 15 seconds or past the generated length, it trims - silently and deliberately, so a long ref clip "losing its tail" is expected behaviour.
• This is a conditioning node, not an output node: its latent feeds a Sampler (run it through the pack's activation-chunk node on low-VRAM cards), and refine_context feeds the face-refine node. Skip that pairing and you're leaving the pack's best trick unused.