Hunyuan 3 Loader (INT8 Budget)

If you've got a 96GB-class card (RTX 6000 Pro Blackwell, H100, that neighborhood), Hunyuan 3 Loader (INT8 Budget) (class HunyuanImage3Int8LoaderBudget) is the loader that gets you ~98% of full-precision quality from Tencent's 80B model without the 160GB of BF16 weights. The tradeoff: INT8 is ~82GB of model, and it realistically wants that 96GB card. The README is blunt about it - forcing CPU offload on smaller cards makes each step take minutes because quantized tensors stream over PCIe all generation long. This is a "big GPU only" node, and it says so up front.

You can also think of it as the grown-up step up from NF4. NF4 gets you onto 24–48GB cards; INT8 is what you run when you can afford the better half of the compromise. The author ships pre-quantized HunyuanImage-3-INT8-v2 weights on Hugging Face, so you don't need to quantize anything yourself.

How it works

The "Budget" in the name is the mechanism. The loader auto-sizes how much of the ~82GB of weights fits on the GPU and spills the rest to CPU RAM, based on two numbers you control:

• reserve_memory_gb (default 20, range 5–80) - VRAM kept free for inference overhead, not the weights. The tooltip says INT8 needs ~20GB minimum; the README's INT8 workflow guidance suggests ~20GB for 1.5–2MP work and more if you're targeting 4K+, so Smart mode has headroom without forcing offload.
• gpu_memory_target_gb (default 80, range 10–128) - how much VRAM the INT8 weights are allowed to occupy. 80GB is the sweet spot on a 96GB card (96 − 12 auto-inference-reserve − 4 slack ≈ 80). Lower it and more weight spills to RAM.

Then there's model_name (auto-detected from your ComfyUI/models/ folder - it expects a folder named HunyuanImage-3-INT8), force_reload (skip the cache), and an optional unload_signal input you can drive from an unload node so the loader loads after memory is freed. The output is a HUNYUAN_MODEL object for any compatible generate node.

How to install it

Install the pack once:

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

Restart ComfyUI and hard-refresh the browser. Then get the weights - the README recommends the v2 pre-quantized INT8 checkpoint:

cd ComfyUI/models
huggingface-cli download EricRollei/HunyuanImage-3-INT8-v2 --local-dir HunyuanImage-3-INT8-v2

The INT8 model is ~85GB on disk and wants 128GB+ system RAM when it has to offload.

Common issues & troubleshooting

Match the loader to the right generate node. The README's compatibility table is emphatic: INT8 Budget pairs with Hunyuan 3 Generate (Large Budget) (or the standard Generate on a 96GB card with the model fully resident), and you should keep offload_mode off for ≤2MP work or you'll pay CPU-speed steps for no reason. Do not use it with the base Large/Offload node's forced CPU offload - the quantized weights don't move to CPU correctly.

OOM at load. Lower gpu_memory_target_gb so more weight spills to RAM, or raise reserve_memory_gb if the failure is at inference. If you're under 80GB VRAM, honestly consider NF4 instead - INT8 on a 48GB card is a study in patience.

VRAM "mysteriously" full. The loader has an unload_signal input for a reason - wire a Force Unload node's output into it when cross-tab pollution is biting you.