Advanced MiniMax H3 Image to Video (Combined)

The plain UC_AdvancedMiniMaxH3ImageToVideo builds your H3 conditioning and latent. This is the "Combined" sibling that goes one step further: it also takes the diffusion model and returns it back, patched when needed - so a single node can hand you model, positive, and LATENT, all three wires for the sampling branch, from one place. It's the node to reach for when your H3 workflow wants reference-derived first/last keyframes and you'd rather not assemble the patching yourself.

What "Combined" adds

The core schema matches the Advanced node - clip (H3's Qwen3-VL 32B encoder, post UC_MiniMaxH3ClipProjectionPatcher), vae, prompt, width/height (1344×768 defaults), length (frame count snapped upward to H3's 17k+5 grid), multiplier, ref_image_size, vlm_resolution - plus the same optional first_frame, last_frame, visual_fusion_config, reference_images, fusion_images, and media_config.

The difference is in the ref_image_size combo and the new model input/output. The Advanced node's match/max/none behavior is retained, and "first" modes are added: first routes the first ordered reference to frame zero, first + last also routes the final reference to the final frame when more than one exists, and match/max sizing then applies to the remaining native references. Each endpoint gets VAE-encoded once.

The model input is where the magic (and the one real gotcha) lives. In mixed keyframe/reference modes, the node returns a patched clone of the model - the tooltip says it plainly: "Mixed keyframe/reference modes return a patched clone; single conditioning modes return this model unchanged." So when you wire the model output into your sampler, you must use the output, not the original model you fed in. The description hammers this home: "Connect the returned model to the same sampling branch as the positive conditioning and latent." People who wire the original model back and wonder why their keyframes got ignored have found the node's one trap.

The setup you can't skip

Same as the sibling: load the encoder via Core's Load CLIP with type minimax, patch it with UC_MiniMaxH3ClipProjectionPatcher, and drop a matching Qwen3-VL→H3 projection (4B or 8B, plain or residual-MLP) into ComfyUI/models/clip_projections/. The H3 video VAE is required, and the generic ref_latent_mode must stay off - H3 uses Core's dedicated reference conditioning.

Outputs

Three sockets: model, positive (CONDITIONING), and LATENT. All three go to the sampler. If you don't need keyframe derivation or model patching, the plain Advanced node is simpler and does the same conditioning job; this node exists for the workflow that wants first/last-frame anchors derived automatically from its reference list.

Install

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

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

Restart. Deps are opencv-python and typing-extensions; the real cost is H3's weights (~42.5GB) and remembering the H3 Community License geofences out the US, EU, UK and South Korea for local-weight use.