Load LoRA Stack INT8

This is Load LoRA INT8 for people running more than one or two LoRAs at once - up to 10 slots in a single node instead of chaining that many loaders. But it's not just tidiness. In Stochastic mode, stacking through this node genuinely produces a better result than chaining single loaders would.

How it works

Same three modes as the single-LoRA node: Stochastic, Dynamic, Standard. The reason to actually reach for the stack version specifically is what happens in Stochastic mode - compatible LoRAs get combined first, and the combined delta gets applied to the INT8 weights with one stochastic-rounding pass, rather than one rounding pass per LoRA. Rounding is where you lose a sliver of precision every time it happens, so combining before rounding once beats rounding once per LoRA in a chain. If you're layering three, four, five LoRAs on an INT8 model, that difference compounds.

Like the single loader, this node expects to sit after your model is already INT8 - coming out of Enable INT8 on MODEL or Load Diffusion Model INT8 (W8A8). It's not a replacement for stock Load LoRA before quantization; it's specifically for LoRAs you want applied to the already-quantized model.

The inputs and outputs that matter

• mode - Stochastic, Dynamic, or Standard, same meanings as the single-LoRA node.
• model - required, your INT8 model.
• lora_1 through lora_10, each paired with its own strength_1 through strength_10 (default 1, range -10 to 10) - all optional, all default to None. Only fill in the slots you're actually using.

Output is a single MODEL, ready for a sampler.

How to install it

• ComfyUI Manager - search "ComfyUI-INT8-Fast-Fork", install, restart.
• Manual - cd ComfyUI/custom_nodes && git clone https://github.com/SparknightLLC/ComfyUI-INT8-Fast-Fork, then restart ComfyUI.

No model downloads for this node itself - just your LoRA files in the usual loras folder and the same GPU/PyTorch requirements as the rest of the pack (a recent ComfyUI, an NVIDIA GPU with real INT8 throughput, a matching PyTorch build).

Common issues & troubleshooting

Only using one or two LoRAs? Just use Load LoRA INT8 instead - it does the same job with a simpler node, and you're not losing the stacking benefit until you're combining three or more.

Leave unused slots as None. All ten lora_N inputs default to None, so there's no need to fill every slot - set only as many as you're actually stacking.

Getting weaker or noisier results than expected with several LoRAs. Make sure you're using Stochastic mode. That's the mode where stacking actually pays off - the node combines the compatible LoRAs before rounding once, instead of rounding after each one. Standard mode skips INT8-aware handling entirely and won't give you that benefit.

Wrong place in the graph. Same rule as everywhere else in this pack: this node goes after the model is already INT8, not before. If you need LoRAs baked in before quantization, use stock Load LoRA nodes ahead of Enable INT8 on MODEL with bake_loaded_loras on, then use this node for anything you want to add afterward.