Adv MiniMax H3 Image to Video (Temporal TokenFusion)

This is the sibling to the Temporal Fusion node in the same pack, and the difference between them is one word in the name: Token. Where Temporal Fusion encodes every extra video sample lane separately with Qwen3-VL and then stitches the results together in the finished conditioning, this variant merges the video features before the encode runs - token-level fusion that needs only one Qwen pass per schedule. Same temporal density, same "ordinary video token budget" contract, roughly one big encode instead of one per lane.

Why you'd pick this one. If you've read the Fusion article you know the pitch: MiniMax H3's video encoder only sees a sample of frames, so fast motion between the samples gets invented, and the pack's temporal-density experiment fixes that by sampling offset frame "lanes" and fusing them back into the fixed token budget. The Fusion node pays for that with a full Qwen encode per lane. On the 32B Qwen3-VL text encoder that MiniMax H3 conditions through, encode passes aren't free - they're seconds and VRAM you feel. TokenFusion fuses the corresponding video features - including their DeepStack vision layers - first, and hands the merged token sequence to a single Qwen encode. If you're GPU-tight, this is the temporal node you want to try before its sibling.

Mechanically it's still grounded in the same experimental frame-sampling machinery: you set temporal_density and the consensus/spatial method on the MiniMax H3 Media Configurator (UC_MiniMaxH3MediaConfig), and density 1 is the "act like ordinary sampling" escape hatch. The configurator controls Picture timestamp formatting, video_fps, and motion guidance too, but the density field is the one this node exists for.

Inputs worth knowing.

• clip - the Qwen3-VL 32B MiniMax H3 text encoder (qwen3vl_32b), loaded with Core's Load CLIP using type minimax. Wrong encoder, immediate error.
• media_config - temporal density and fusion method live here; this is the switch that turns the experiment on.
• video - the complete 24 fps frame batch the temporal lanes are drawn from.
• length - frame count, snapped up to H3's 17k+5 temporal grid at 24 fps.
• text_blend_config - consensus weighting for the merge (disconnected uses a default).
• multiplier - scales the final conditioning; leave at 1.0 unless you know why you're touching it.

Frame references go through first_frame/last_frame (VAE anchors plus the first two numbered Qwen pictures) or ordered reference_images - never both, and never combined with a reference video. vlm_video_resolution trades visual tokens against encode cost; higher captures more per-frame detail.

Outputs. positive CONDITIONING for the sampler's positive socket, and LATENT for the KSampler - the standard advanced-H3 hand-off, then decode and save as usual.

Install. It ships in the same pack as its sibling, so the steps are identical: ComfyUI Manager → search ComfyUI-UtilsCollection, or

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

then restart. The pack's own deps are light (opencv-python, typing-extensions, unifiedefficientloader); the cost is the model stack - MiniMax H3 checkpoint (~42.5 GB), Qwen3-VL 32B encoder, H3 VAE - which nothing downloads for you. Same licence warning applies: the MiniMax H3 Community License excludes the EU, UK, South Korea and the US from its applicable territory.

Reality check. Both temporal nodes are experiments with zero community mileage behind them - treat results as research-grade. If density 1 doesn't reproduce your stock workflow's output, you've found a bug worth reporting rather than a quirk to route around. If you just need H3 image-to-video and your motion is fine, none of this is for you; it's for the shots where the smearing between frames is the problem and you don't want to pay a larger conditioning to fix it.