Trellis_multiimage_loader

Nobody searches for this node, and its search stats agree - but it's the one that fixes TRELLIS's biggest weakness. Give the pipeline a single image and it has to invent the back of the object, which is where you get those melted, guesswork silhouettes. Give it two or three views of the same thing and the reconstruction stops guessing. This tiny node is how you do that in this pack, and skipping it is the most common reason people think TRELLIS is worse than it is.

What it does

Pure plumbing, but useful plumbing. image_a is required; image_b and image_c are optional. It center-crops or pads each image to a square at the largest height, resizes them all to that common size, then concatenates them into a single batched tensor. That's the whole job - a batch-prep node you'd normally have to hand-roll.

Why it matters

Wire its IMAGE output into Trellis_Sampler's image input, flip multi_image to true, and when the batch is a multiple of three the sampler splits it into triplets and calls the pipeline's true multi-view path with your chosen multiimage_algo - multidiffusion or stochastic. The classic setup is three angles of the same object: front, three-quarter, back. Same object, different views; the reconstruction fuses them into one coherent asset instead of hallucinating the hidden side.

Inputs and output

• image_a (required) - the first view.
• image_b, image_c (optional) - the second and third views. Any one you leave unplugged just shrinks the batch; two views still beats one.

Output is a single IMAGE tensor (a batch of the up-to-three views), sized to the tallest input.

Gotchas

• The images should be different angles of the same object. Different objects in one batch just confuse the sampler.
• Multi-image mode only fires when the batch count is divisible by three, and only when the sampler's multi_image toggle is on. Miss either and this node is a glorified batch concatenator.
• Every image gets center-cropped to a square, so compose your shots accordingly - a wide shot will lose its edges.

Install

There's nothing to install beyond the pack itself. This node is part of smthemex/ComfyUI_TRELLIS, so the git clone https://github.com/smthemex/ComfyUI_TRELLIS.git into ComfyUI/custom_nodes (plus the pack's compiled-extension install - see the Trellis_LoadModel page) brings it along with zero extra dependencies. If the sampler works, this node works.