Hunyuan 3 Loader (88GB GPU Optimized)

Straight answer first: Hunyuan 3 Loader (88GB GPU Optimized) (class HunyuanImage3SingleGPU88GB) is deprecated, and the pack's own README marks it with a skull-and-crossbones: "❌ DEPRECATED | Everything | Use Full BF16 Loader instead." If you're starting a new workflow today, don't use this node. Use the Full BF16 loader (or one of the quantized loaders) and save yourself the maintenance headache.

So why does this page exist? Because it's still in the pack, old workflows still load it, and if you open a workflow from November 2025 you may well see this node staring back at you. It's useful to know what it did, why it's gone, and what to swap it for.

What it was

This was the single-GPU BF16 loader for the era before the pack's Full BF16 Loader (GPU) arrived - "88GB" because it was tuned for big single cards in the 80–96GB range. It loaded the full-precision HunyuanImage-3.0 model onto one GPU and gave you a single meaningful knob:

• reserve_memory_gb (default 14, range 8–24) - how much VRAM to hold back for inference. The old configuration guide suggested ~14GB to leave room for VAE decode and activations while the ~80GB of BF16 weights sat on the card.

Its only other input is model_name (expects the full HunyuanImage-3 folder in ComfyUI/models/), and it outputs a HUNYUAN_MODEL for the generate nodes.

Why it's deprecated

The Full BF16 Loader(s) that replaced it do the same job better - cleaner memory control, a target_resolution dropdown that reserves the right amount of inference headroom per job (35GB for Auto, 8GB for 1MP fast, 75GB for 4MP+), and clear_vram_before_load for the cross-tab pollution problem. Keeping two near-identical BF16 loaders around was pure duplication, so the author retired this one. It still exists for backward compatibility; it just doesn't get bug fixes or new features.

How to install it

It's in the pack either way:

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

Restart ComfyUI. Or search HunyuanImage in ComfyUI Manager. The node will be there - you just shouldn't build on it.

Common issues & troubleshooting

Your saved workflow uses it and you're worried. Don't be. Delete the node and add Hunyuan 3 Loader (Full BF16) (or Full BF16 GPU) with the same model_name folder - the HUNYUAN_MODEL output is identical, so your generate nodes re-connect without changes.

It OOMs at load. The old default of 14GB reserve was tuned for ~1MP generation. For larger outputs you'd raise reserve_memory_gb, which is exactly the fiddly manual tuning the replacement's target_resolution dropdown automates.

One general truth about full BF16 either way: it's ~160GB on disk, ~80GB of VRAM, and the README wants 80GB+ VRAM or a 192GB+ RAM rig to offload into. If you don't have that hardware, the NF4 or INT8 quantized loaders in this pack are the ones actually worth your time.