Hunyuan3Dv2Conditioning

Hunyuan3D-2 is Tencent's 3D generation model - the turbo, mini, and multiview variants - and unlike image models, it doesn't take a text prompt. Its conditioning is a picture. Hunyuan3Dv2Conditioning is the node that turns a CLIP vision encoding of your reference image into the positive/negative conditioning the 3D sampler consumes. It's been in core since March 2025, when the initial Hunyuan3Dv2 implementation landed.

How it works

One input, two outputs, and the whole mechanism is visible in the source:

• clip_vision_output - the output of a CLIPVisionEncode node (which itself takes a CLIPVisionLoader-loaded vision model plus your image).

The node grabs the vision encoder's last hidden state and uses it directly as the positive conditioning. The negative is the same shape, filled with zeros. That's it. No pooled projection, no text fusion, no merging - the 3D model is conditioned by the image's visual features alone.

Outputs

• positive - the image embeddings, ready for the KSampler.
• negative - zeros. This is the "generate nothing" reference that classifier-free guidance uses to define contrast.

Building the graph

The chain is: load a CLIP vision model → CLIPVisionEncode on your source image → this node → KSampler with the Hunyuan3D-2 model → 3D decode. For a single-view generation, that's the whole path. And there's a sibling worth knowing: Hunyuan3Dv2ConditioningMultiView, which takes up to four CLIP_VISION_OUTPUTs (front, left, back, right) and conditions on all of them - that's the one for reconstruction from multiple views, and it's the difference between "make a 3D shape from this photo" and "reconstruct the full object."

The gotchas

The "negative" being literal zeros can surprise people who come from image models with negative prompts - there's no text here at all. If your 3D output looks ungrounded, it's your reference image quality, not your prompt. Second: this node consumes CLIP_VISION_OUTPUT, the processed output of a CLIPVisionEncode node, not an IMAGE and not a raw CLIP_VISION model. Wiring the vision model in directly is the classic connection mistake - the slot types won't even match in a normal graph, but with mismatched workflows you can end up feeding the wrong thing through a converter and getting a subtle failure.

Third, resolution and framing matter disproportionately in 3D. A centered, unobstructed reference gives the model what it needs to infer the back side; a portrait crop or heavy shadowing gives it nothing, and you get a blob. The model family itself is solid and the community consensus is that Hunyuan 3D is where local 3D generation actually got usable - this node is just the small, exact piece that makes the pipeline go. Ships with ComfyUI core, no install.