Advanced MiniMax H3 Image to Video

MiniMax H3 is the 33B omni-modal video model that dropped in August 2026 with native audio, and its ComfyUI pipeline is demanding: it needs a Qwen3-VL text encoder projected into H3's 32B conditioning space, a separate video VAE for keyframes, reference images that become both VLM pictures and VAE anchors, and a joint video/audio latent. Getting all that wired by hand is a genuinely unpleasant afternoon. UC_AdvancedMiniMaxH3ImageToVideo is the pack's node that collapses most of it into one box: feed it the encoder, the VAE, a prompt, and images, and it returns both the positive conditioning and the matching LATENT ready for sampling.

Setting it up (the part people skip and then blame the node)

The README is explicit about the two non-obvious setup steps, and skipping either produces "why is this garbage" hours:

1. Load the encoder with Core's Load CLIP using type minimax, then run it through UC_MiniMaxH3ClipProjectionPatcher, which projects a Qwen3-VL 4B or 8B encoder into H3's conditioning space.
2. Download one matching projection file into ComfyUI/models/clip_projections/ - e.g. mmh3-4b-ClipProj-v3.1.safetensors or the 8B equivalent, plain or residual-MLP variants - from the linked NicoLab28 repo. Only one is used at a time.

This node then takes the patched CLIP as its clip input plus the H3 video vae, and handles the rest.

The inputs that matter

• clip - the H3 Qwen3-VL 32B text encoder (post-projection patcher).
• vae - MiniMax H3 video VAE, used for frame anchors and native references.
• prompt - raw H3 prompt. Picture labels are supplied by Core from the connected image roles, so don't hand-write <Picture N>.
• width / height - 1344×768 defaults, 32-step.
• length - frame count at 24 fps (default 124). The tooltip's key detail: it's snapped upward to H3's 17k+5 temporal grid, so don't fight it with odd counts.
• ref_image_size - match (limit native references to the generation pixel area), max (short edge capped at 2048), or none (images stay Qwen pictures only, no native VAE reference conditioning). The none mode is how the README's eight-image, 1024-VLM workflow runs - visual-token conditioning without VAE reference encoding.
• vlm_resolution - equivalent-square Qwen3-VL target (256–3584, default 384), independent of VAE sizing.

Optionals are where it gets interesting: first_frame/last_frame become VAE anchors and numbered Qwen pictures; reference_images are ordered native H3 references (mutually exclusive with explicit frames); fusion_images and visual_fusion_config handle Qwen-only fusion routing; and media_config (from UC_MiniMaxH3MediaConfig) adds Qwen-only Video timeline images with timestamps plus optional native audio via an H3 audio VAE.

Outputs

positive (CONDITIONING) and LATENT. Wire both into the same sampler - this node is a conditioning and latent factory, not just an encoder. The README's worked example reproduced a 12.25-second sequence's framing in seven sampling steps on a 16GB GPU, which is the right mental model for what this node is trying to make easy.

Install

ComfyUI Manager → search ComfyUI-UtilsCollection → install, or clone:

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

Restart. Python deps are light (opencv-python, typing-extensions); the heavy lifting is the H3 weights themselves, which is a big download and - important - the H3 Community License excludes the US, EU, UK and South Korea from local-weight use, so check that before you spend the bandwidth. One more thing: the pack's generic ref_latent_mode must stay off here; H3 uses Core's dedicated reference conditioning instead.