MiniMax H3 Media Configurator

H3 is an omni-modal model: images, video and audio all enter as one conditioning context. UC_MiniMaxH3MediaConfig exploits that. It takes the Picture slots your H3 conditioning already has and anchors each one at a formatted timestamp - so instead of "here are some images, do something," you're saying "image A happens at 0.5s, image B at 2.0s, image C at 4.5s." It can also attach native audio conditioning, which is the part no video model without H3's unified architecture could even offer.

It's part of silveroxides/ComfyUI-UtilsCollection, and it's an input-node: it produces a media_config object consumed by the pack's Advanced MiniMax H3 encoder nodes (UC_AdvancedMiniMaxH3ImageToVideo and friends). You don't use it standalone; you use it to make those encoder nodes timeline-aware. The README's headline example is an 8-image storyboard workflow where each chronological frame is associated with a timestamp, and it reproduces the framing and progression of a 12-second source sequence in seven sampling steps on a 16GB GPU.

What's on the node

• timestamps - sequential timestamps for your existing Picture slots. Accepts a list, or comma/semicolon/newline-delimited values in seconds or HH:MM:SS format. Count should line up with your pictures.
• timestamp_format (default 0.0s) - how the timestamp reads when expanded into the structure text.
• structure - a per-shot template using the placeholders <<time>>, <<picture>>, <<visual>>, and <<shot>>. Default: "At <<time>>, <<picture>>: <<visual>> (from <<shot>>) is fully anchored." Edit it to change how the timeline instruction is phrased to the VLM.
• audio (optional) - native H3 reference audio. Qwen receives only an "Audio" label; hard sync with Video timestamps isn't guaranteed, per the tooltip.
• audio_vae (optional, required with audio) - the MiniMax H3 audio VAE, which resamples the audio to the model's native rate and creates reference audio rows.
• media_config output - the object for the Advanced encoder nodes.

Installing it

Via the pack:

• ComfyUI Manager: search "ComfyUI-UtilsCollection".
• Or:

cd ComfyUI/custom_nodes
git clone https://github.com/silveroxides/ComfyUI-UtilsCollection

Restart ComfyUI. The node itself needs no extra downloads, but you're building on H3 - big models, territory-restricted license (US, EU, UK, Korea excluded), and if you want the audio path you'll also need the H3 audio VAE.

Gotchas

Two expectations to set. First, timestamps are instructions to the VLM, not hard frame anchors - the README says the results remain stochastic, meaning the model tries to honor your timeline rather than cut on exact frames. Second, the audio is native reference audio, not a synchronized soundtrack: Qwen just sees an Audio label, so don't expect lips to match. The structure template is where a lot of tuning happens; if the model ignores your timing, rephrasing the template usually does more than moving timestamps around. And keep timestamps in order - the node expects sequential input, so a scrambled list makes for scrambled storytelling.