Hunyuan3D V1 - Text2Image Pipeline Load

This node's whole job is to load the Hunyuan DiT text-to-image model so the Text2Image node can draw your prompt. It's a loader with one input and one output - about as thin as nodes get. But that one input, save_memory, is the difference between the text-to-3D path fitting on your card and silently dying, so it's worth understanding.

What save_memory does

When save_memory is off (the default), the pipeline loads Hunyuan DiT onto the GPU and keeps it there for the whole session. Fast, but the model's VRAM stays reserved the entire time, stacked on top of the multi-view and mesh models that come later in the chain.

Turn it on and the node gets lazy in the smart way: it keeps the pipeline on CPU, shuttles it to cuda:0 only while an image is actually being generated, then moves it back off and frees the memory between runs. Generation gets slower - each run pays a transfer cost - but the big text-to-image weights don't squat in VRAM for the entire workflow. If you're trying to run the whole Hunyuan3D-1 stack (which is genuinely heavy) on a modest card, this is the knob that makes the text-to-3D route viable.

How the load itself works

The node loads the distilled Hunyuan DiT - Tencent-Hunyuan/HunyuanDiT-v1.1-Diffusers-Distilled - from the pack's weights/hunyuanDiT folder, via the diffusers AutoPipelineForText2Image loader with perturbed attention guidance enabled. That's why the Text2Image node defaults to 20 steps rather than 50: it's a distilled model. All of that setup is invisible; you just see the pipeline_config output, which you wire straight into the Text2Image node's pipeline_config input.

The practical take

There's one decision here, and it's a memory-versus-speed tradeoff. Have the VRAM? Leave save_memory off and enjoy faster runs. Struggling? Flip it on - it's the single easiest way to shave the text-to-3D path's footprint without touching quality.

Install

This is the pack-wide install, and it's not pleasant: fresh ComfyUI recommended, Python 3.12, C++ build tools, pytorch3d from source, an Open3D py3.12 dev wheel, nvdiffrast, Ninja, plus the pack's requirements.txt. Note this node specifically needs the diffusers weights, which are a separate download from the 3D weights:

huggingface-cli download Tencent-Hunyuan/HunyuanDiT-v1.1-Diffusers-Distilled --local-dir ./weights/hunyuanDiT

If you'd rather skip this node entirely, the README notes a native ComfyUI route: drop hunyuan_dit_1.2.safetensors into ComfyUI/models/checkpoints and run HunyuanDiT with core ComfyUI nodes. But if you're on the wrapper's text-to-3D path, this loader is the price of entry.