MiniMax H3 • Long Media Setup

This is where every long H3 run starts. Long Media Setup is the orchestrator: it takes your prompt, duration, references and workflow mode, builds the multi-segment plan, encodes the source media into the AV latent, and hands the whole thing to the sampler. If you're new to this pack, this is the node you'll stare at longest, so it's worth understanding the handful of settings that actually change your output.

The core inputs

• clip, vae, audio_vae - the H3 text encoder, video VAE and audio VAE. Three separate model files; the audio VAE is easy to forget and the pack needs it even for silent video.
• prompt - the text. There's also a direct STRING socket if you want to drive it dynamically.
• width, height, video_fps (24) - canvas and frame rate.
• manual_duration (10 s) and duration_source (auto / manual / audio / video / longest_input) - where the final length comes from. Pick audio and a connected reference and the clip lengths itself to the soundtrack.
• segment_seconds (5) - the amount of new visible timeline per segment. The 22-frame continuation overlap is extra hidden context and does not count against this. This is your main VRAM lever: smaller segments = lower peak memory.
• overlap_frames (22) - continuation context between segments, snapped to the H3 grid.
• reference_budget (low/medium/high/max) - how aggressively reference payloads are limited. On constrained cards start at low; big Picture/Video/Audio references balloon the packed sequence and can dominate VRAM.
• audio_mode - auto (legacy), preserve (restore source audio at output), generate (H3 makes the audio), reference_only (input as reference, output generated), preserve_reference (input as timing reference, source restored), and lip_sync (source audio drives timing, untouched track restored). If you're dubbing or matching a soundtrack, this is the input.

Workflow modes - the ones that matter

workflow_mode is the decision that shapes everything:

• hybrid_auto (default) - image_1 is the opening frame anchor; connect image_2 and it becomes the final frame anchor; the rest are Picture references.
• segmented_continuation - auto-splits one continuous prompt into equal fixed segments. Good for VRAM control and stability with a single prompt. It does not map different actions to different time points - that's what multiclip is for.
• multiclip - uses the Long Media Planner; every clip gets its own prompt, duration and optional seed. This is the one for "different shots at different times."
• ref2va_full - all images are plain <Picture N> references, no anchors.
• video_ref_edit - video_1 carries motion/camera/composition, image_1..9 do identity/style replacement. If the source video has a soundtrack, load and connect it to audio_1 separately - an IMAGE connection never carries audio.
• loop - image_1 becomes both first and last frame for seam-friendly loops.

One trap the README calls out: if you connect a clip_plan (Planner) while workflow_mode is anything other than multiclip, Setup silently ignores it. The Planner is authoritative only in multiclip, and that's deliberate so a leftover connected Planner can't override your mode.

Outputs

positive (CONDITIONING) and long_media_av (LATENT) into the sampler, long_media_plan (the plan object) into both Sampler and Decode, plus duration_seconds, passes and a report. If release_guard is off, the console prints the full internal diagnostics - handy for debugging, noisy otherwise.

Install

cd ComfyUI/custom_nodes
git clone https://github.com/vizart-vj/ComfyUI-MiniMax-H3-LongMedia

(Manager search "MiniMax-H3-LongMedia" also works), restart, and load the H3 weights from MiniMaxAI/MiniMax-H3 - checkpoint + both VAEs, ~42.5 GB, under the H3 Community License that excludes the US, EU, UK and Korea. Check your region before downloading.

Troubleshooting

If your references seem to do nothing, check the mode: in ref2va_full nothing is anchored, and in hybrid_auto only image_1/2 get anchor status - a connected image in the wrong slot reads as a plain reference, not a keyframe. If runs OOM, cut segment_seconds, set reference_budget=low, and leave Dynamic VRAM enabled (the pack explicitly wants you to not launch with --disable-dynamic-vram). And if passes comes back unexpectedly large, you've set manual_duration way above segment_seconds - which is fine, but it's worth knowing segmentation is doing the work, not single-pass generation. Community wisdom on H3 is that it's genuinely good but genuinely heavy; the people running it on 8–12 GB cards are the ones leaning on exactly these segment and budget knobs.