Hunyuan 3 Loader (Full BF16 GPU)

"Hunyuan 3 Loader (Full BF16 GPU)" is the single-GPU, full-precision loader - the version of the BF16 loader that gives you explicit memory control instead of the resolution-based reservation dropdown. If you have one 80GB+ card and you want to tell the loader exactly how much VRAM it's allowed to use, this is the node. The full-precision 80B HunyuanImage-3.0 is the reason this node exists, and it makes no compromises: no quantization, full bfloat16 quality.

What's different from the plain Full loader

Where the base Hunyuan 3 Loader (Full BF16) reserves VRAM through a target_resolution dropdown (35GB auto, up to 75GB for 4MP+), this one exposes the mechanics directly:

• gpu_device - which CUDA device to load onto (the enum is just cuda:0; this is a one-GPU node by design).
• max_gpu_fraction - the fraction of GPU memory the model is allowed to occupy, default 0.9. Crank this down to give inference and the VAE more breathing room.
• reserve_memory_gb - explicit VRAM headroom left free for inference, default 8GB. This is the knob to raise when you're rendering large images or seeing OOM at generation time rather than load time.

Add force_reload? No - this loader is leaner than the others: just those controls, plus a HUNYUAN_MODEL output to feed any base-model generate node. It doesn't even take an unload_signal input. It's purpose-built for the "I know exactly what my card can hold" crowd.

The model folder it wants is the full BF16 release in ComfyUI/models/:

cd ComfyUI/models
huggingface-cli download tencent/HunyuanImage-3.0 --local-dir HunyuanImage-3

When to pick this over the Full loader

Honestly, the two overlap a lot. This node is for people who want the reserve slider without the resolution dropdown deciding for them - maybe you're rendering a fixed resolution and want every free gigabyte on the weights, or you're sharing the GPU and want to hard-cap the model's footprint. The README's own positioning calls the plain Full loader the general recommendation and this one the "single GPU with reserve slider" option. If you don't know which you want, start with the plain Full loader's Auto preset and only move here when you're tuning.

Installing

Same pack as everything else:

cd ComfyUI/custom_nodes
git clone https://github.com/EricRollei/Comfy_HunyuanImage3
cd Comfy_HunyuanImage3
pip install -r requirements.txt

Restart ComfyUI, or install through ComfyUI Manager (search "Comfy_HunyuanImage3"). Requirements include torch>=2.8.0, transformers>=4.47, and accelerate>=1.2.1.

The reality check

This node inherits all the requirements and none of the excuses of full BF16: 80GB+ of VRAM ideally, ~160GB on disk, 192GB+ of system RAM if you're running it on anything less. And because it keeps everything on one GPU, it's the wrong tool if your plan is "two 4090s and some hope" - that's what the Multi-GPU BF16 loader is for. If a single card can't hold the full model comfortably, step down to NF4 and don't look back. The license note applies here too: the pack's code is CC BY-NC 4.0 (non-commercial) even though Tencent's model weights are Apache 2.0.