Adv MiniMax H3 Image to Video Combined (TokenFusion)

If you read the TokenFusion article first, this is the sibling that changes one thing that matters: it takes the model as input and gives it back to you patched. UC_AdvMiniMaxH3ImageToVideoTokenFusion returns conditioning and a latent and expects your model to be wired in separately. This "Combined" variant returns model (a patched clone when it needs one), positive, and LATENT - and all three are meant to feed the exact same sampling branch. That's the "combined" contract: model, conditioning, and latent travel together, which is how the pack's MiniMax H3 nodes work best.

The other thing "Combined" adds is the keyframe trick. The plain node treats your images as Qwen pictures - tokens the VLM reasons about. This one can go further via ref_image_size: set it to one of the "First" modes and the first ordered reference gets routed to frame zero as a real VAE-encoded keyframe; use "first + last" and the final reference anchors the final frame too. That's meaningful for video because a true keyframe at the endpoint constrains the start and end frames at the latent level, not just as a prompt hint. Non-endpoint references still get treated as native H3 references and numbered Qwen pictures, and each endpoint gets VAE-encoded exactly once. Frames and references still can't be mixed with explicit frame inputs in reference mode, and a mixed keyframe/reference setup returns the patched model clone while single-conditioning modes pass the model through unchanged.

Inputs otherwise mirror the TokenFusion node: clip is MiniMax H3's Qwen3-VL 32B encoder (qwen3vl_32b), vae is the H3 video VAE, prompt is the raw prompt with Picture labels supplied by Core from connected image roles. width/height default to 1344x768, length defaults to 124 frames (snapped upward to H3's 17k+5 temporal grid - don't be surprised when 124 becomes a nearby number), multiplier scales the Qwen conditioning and pooled output without touching VAE keyframes or the latent, and vlm_resolution (default 384) sets the equivalent-square resolution the encoder sees.

The optional autogrow sockets - first_frame, last_frame, reference_images, fusion_images - plus visual_fusion_config and media_config form the full routing contract, and there are two subtleties worth flagging. In reference mode, fusion_images pairs with reference pictures by index (one image on fusion_image_1 broadcasts to all of them), and the media_config can add Qwen-only Video blocks and optional native audio conditioning - but video blocks are never fusion targets. If a socket doesn't match the expected layout, you get an error rather than silent weirdness, which is honestly the friendlier failure mode.

Install and model requirements are identical to the rest of the pack and covered in the TokenFusion article: Manager or git clone, light Python deps, then the real work of grabbing MiniMax H3, the Qwen3-VL 32B encoder (or a 4B/8B plus a clip projection into ComfyUI/models/clip_projections/), and the H3 VAE. Same license caveat applies - H3's community license excludes the US, EU, UK, and Korea, so check your territory before downloading anything.

Where people get burned with this node is almost never the node itself - it's wiring the three outputs into different samplers or forgetting that keyframe mode and explicit frame inputs are mutually exclusive. Keep model, positive, and LATENT on one branch and you're most of the way there.