Adv MiniMax H3 Image to Video (TokenFusion)

MiniMax H3 is the big one for 2026 video: MiniMax's 33B omni-modal model (the open-weights version of Hailuo 3.0) that takes text, images, video, and audio as one context and generates clips with native, jointly-produced sound. ComfyUI had day-0 support, and this pack is where a lot of the day-0 workflows came from - silveroxides is the same person behind the Chroma GGUF quants and ComfyUI_SamplingUtils, i.e. a power user who lives on the bleeding edge of VLM conditioning rather than someone shipping Yet Another Upscaler.

So what does this node do? It's a variant of the pack's Advanced MiniMax H3 Image to Video encoder, and the "TokenFusion" in the name is the whole point. Instead of encoding each reference image separately and blending the results afterwards, it fuses the visual tokens from your frame, reference, and fusion images first, then runs a single shared Qwen3-VL encode. One encode means fewer passes over the text encoder and - in theory and usually in practice - tighter coordination between the pictures and the prompt, because the model never has to reconcile separately-encoded conditionings.

The node takes a clip (MiniMax H3's Qwen3-VL 32B text encoder, qwen3vl_32b), a vae (the H3 video VAE), and your raw prompt - Picture labels are supplied by Core from whatever image roles you connect, so don't hand-write them. width/height default to 1344x768, length defaults to 124 frames, which is the interesting gotcha: frame count is snapped upward to MiniMax H3's 17k+5 temporal grid, so 124 becomes whatever the grid lands on. multiplier scales the final Qwen conditioning and pooled output without touching the VAE keyframes or latent, and vlm_resolution (default 384) controls the equivalent-square resolution the Qwen3-VL encoder sees the images at.

Optional inputs are where the real work happens: first_frame and last_frame become VAE-encoded anchors at frame zero and the end, reference_images (an ordered autogrow list) are native H3 references that also become numbered Qwen pictures, visual_fusion_config adds a spatial fusion method, and media_config can add a Qwen-only Video timeline plus optional native audio conditioning. One rule to remember: reference mode and explicit first/last frame inputs are mutually exclusive - the node will error if you mix them.

You get two outputs - positive (CONDITIONING) and LATENT - and they must both feed the same sampling branch, along with the same model. Forget that and you'll get mysterious inconsistency between what the latent expects and what the conditioning describes.

Installation is the pack-standard: ComfyUI Manager search for "ComfyUI-UtilsCollection", or:

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

Its real dependencies are light (opencv-python, typing-extensions) - the weight is in the models. You need the MiniMax H3 checkpoint, the Qwen3-VL 32B text encoder (Core's Load CLIP with type minimax), and the H3 VAE. If 32B is too heavy, the pack's UC_MiniMaxH3ClipProjectionPatcher lets you use a 4B or 8B Qwen3-VL with a small projection file in ComfyUI/models/clip_projections/ instead. And before you go far: H3's license geofences out the US, EU, UK, and Korea - if you're in one of those, the local weights aren't licensed to you at all, API or nothing.