Hunyuan3Dv2ConditioningMultiView

If you've generated a 3D model from one photo and gotten a front that looks great and a back that looks like the model gave up, Hunyuan3Dv2ConditioningMultiView is the fix. Hunyuan 3D v2 can take visual context from multiple camera angles, and this node is the bridge that turns up to four of them - front, left, back, right - into the conditioning the 3D diffusion model samples from. Give it a full sweep and it reconstructs the whole object instead of hallucinating the hidden side.

This is a conditioning-only node, which surprises people who expect it to do the 3D work. It doesn't generate anything; it packages what your vision models already extracted from the four views.

How it works

The inputs are CLIP vision outputs - the embeddings your CLIP Vision encoder produced for each angle. The node takes whatever of the four views you connected (front/left/back/right are all optional), adds a small sinusoidal positional embedding to each so the model knows which token came from which angle, and concatenates them all into a single sequence along the token dimension. That becomes the positive conditioning; the negative conditioning is the matching shape of zeros - the standard "here's what the object is not" placeholder. So even when you connect only a front view, the node still emits both positive and negative, which is why the outputs are always a pair.

Inputs and outputs

• front / left / back / right - all optional CLIP_VISION_OUTPUT. Connect what you have; a 3/4 sweep already beats a single view.
• Outputs: positive and negative CONDITIONING, wired into the Hunyuan 3D v2 sampling nodes alongside the empty 3D latent.

Where it fits

Ships with ComfyUI core; it's part of the Hunyuan 3D v2 support that landed in the 2026 timeframe, sitting next to its single-view sibling (Hunyuan3Dv2Conditioning) and the 3D VAE/sampler nodes. The actual model weights - the Hunyuan 3D v2 checkpoints, the CLIP vision encoder - go in your ComfyUI models folders. No install.

Common issues

The main one is underfeeding it: one photo is still the most common setup, and Hunyuan 3D v2 will happily invent geometry for the unseen half. Feed it at least three views if you can - the back is the hardest to hallucinate convincingly. Also, the views need to be consistent: a wildly different camera height or exposure between angles produces a conditioning sequence that disagrees with itself, and you'll see it in seams or warped topology. And don't confuse this with the single-view node; the multiview one expects CLIP vision outputs per angle, not raw images - you still need a CLIP Vision encode in front of each input.