Hunyuan 3 Loader (INT8)

"Hunyuan 3 Loader (INT8)" is the middle path through Tencent's HunyuanImage-3.0: 8-bit quantization that keeps ~98% of the full-precision quality while shrinking the 160GB BF16 model to about 85GB on disk. It's the choice when NF4's quality loss bothers you and a full BF16 card isn't in your budget - which, for most people, means it's the choice for a 96GB-class GPU like an RTX 6000 Pro or H100.

Be honest with yourself about what this node does not do. The README is unusually candid: the selective INT8 checkpoints shipped here "realistically need a 96GB-class GPU." You can force it onto a 24GB card with CPU offload, but the quantized tensors stream over PCIe every step and each step takes minutes. The INT8 loader is not a way to make an 80B image model fit consumer hardware - it's a way to get better quality than NF4 on hardware that's already big.

The controls

• model_name - points at the HunyuanImage-3-INT8 folder in ComfyUI/models/. Pre-quantized v2 weights:

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

• force_reload - bust the cache when the loaded state is stale.
• reserve_memory_gb - VRAM left free for inference, default 6GB. The tooltip's guidance: 6GB for standard work, more for >2MP renders since large images eat ~15GB/MP (at which point you'd switch to a Large/Budget generator anyway).
• unload_signal - optional any-type input; wire a trigger here to free the model when downstream is done.

Output is a single HUNYUAN_MODEL that feeds any base-model generate node.

Pairing and performance

The README's recommended setup for INT8 is the INT8 Budget loader + Generate (Large Budget) with offload_mode: disabled - on a 96GB card that combination does ~1MP renders at about 3.9s/step with the model at ~85–95GB of VRAM. Note the "Budget" name: the INT8 loader here is the budget-aware variant with reserve_memory_gb and gpu_memory_target_gb controls baked in (20GB reserve / 80GB target is the suggested starting point), which is why it pairs with the budget generator. The two talk through shared metadata so "smart" offload decisions actually know your card.

Installing

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

Restart ComfyUI. You can also reach it via ComfyUI Manager (search "Comfy_HunyuanImage3"). Requirements include bitsandbytes>=0.48.2, accelerate>=1.2.1, and transformers>=4.47.

The trap

Don't chase INT8 on a card that can't hold it. On 24–32GB cards the correct move is NF4 with the Low VRAM loader/generator pair - the quality delta between NF4 and INT8 is invisible if your render is spending four minutes per step swapping tensors over PCIe. And if you do land on 96GB: for 3MP+ renders even INT8 hits the MoE dispatch_mask memory wall, which is why the pack added the HighRes Efficient generator. INT8 is a quality-over-speed pick, and only on big iron.