Hunyuan 3 Loader (Full BF16)

"Hunyuan 3 Loader (Full BF16)" loads the untouched, uncompressed HunyuanImage-3.0 - all 80 billion parameters in bfloat16. No quantization, no quality trade-off, and no pretending it's a consumer-card thing. This node wants an 80GB+ GPU (A100, H100, RTX 6000 Ada), ~160GB on disk, and if you're below 80GB of VRAM it will happily live in your system RAM - the README recommends 192GB+ of system RAM for that arrangement, and warns that generation gets slow when weights are shuttling between RAM and VRAM.

It's the quality reference point for the whole pack. Every quantized variant - NF4, INT8 - is measured against what this loader produces. If your hardware can genuinely hold it, this is the best-looking HunyuanImage-3.0 you will ever see.

What it actually manages

Three knobs do the interesting work:

• target_resolution - a dropdown that reserves VRAM for inference based on what you plan to render: "Auto (safe default)" reserves a conservative 35GB, "1MP Fast (96GB+)" keeps more weights on GPU for speed at small renders, and it scales up through "4MP+ (2560x1920)" at 75GB reserved. This is the single most important control - reserve too little and the smart offload logic will panic-move weights to RAM mid-run to avoid a crash.
• clear_vram_before_load - when enabled, unloads other models (Flux, SAM2, whatever) from VRAM while keeping a cached Hunyuan around. This is the fix for the "VRAM is mysteriously full" problem you get from cross-tab model pollution.
• force_reload - clear the cache and reload from disk, useful when a failed load leaves orphaned VRAM.

There's also an optional unload_signal input to trigger teardown, and the output is a single HUNYUAN_MODEL.

The weights come from Tencent directly:

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

Pairing

This loader keeps the model in RAM when it doesn't fit on GPU, so it pairs with the Large/Offload generate nodes - not the base Generate node, which assumes a fully-resident model. And for 3MP+ renders on 96GB cards, the HighRes Efficient node is the recommended partner; the standard Large node OOMs at 3MP+ because the MoE's dispatch_mask intermediate tensor grows quadratically with token count, while the efficient variant routes experts with a ~75× smaller footprint.

Installing

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

Restart ComfyUI (or install via ComfyUI Manager, searching "Comfy_HunyuanImage3").

The honest read

If you're here, you already know the punchline: this is the 80B model that the community measured at effectively datacenter-only on release, and the 24GB-card crowd was told to forget it. This loader is the "correct" way to run it when you have the iron. On 24GB cards, "full BF16 with 140GB of weights in RAM" is technically a configuration and practically a stress test - that's what the NF4 loader exists for. Also note the licensing split: the model is Apache 2.0 from Tencent, but the pack's integration code is CC BY-NC 4.0, non-commercial by default.