FLUX LoRA Multi Loader (Legacy - Do Not Use in V2)

If you're on Nunchaku FLUX and want to stack more than one LoRA without chaining a pile of separate LoRA-loader nodes across your graph, this is the node for that. The display name in ComfyUI literally says "Legacy - Do Not Use in V2," which is the author being unusually blunt with you: this is the original (V1) loader, and there's a newer one - FluxLoraMultiLoader_10, "FLUX LoRA Loader V2" - built for ComfyUI's newer Nodes 2.0 / Desktop frontend. If you're starting a workflow from scratch today, go grab that one instead. This node earns its keep only if you've got an existing workflow built on it, or you're on the older ComfyUI 1.x frontend where V2's widget handling doesn't behave.

What it actually does is unglamorous but useful: apply up to ten LoRAs to a Nunchaku-quantized FLUX model in one node instead of ten. lora_count (1–10) decides how many of those slots are actually processed - set it to 3 and slots 4 through 10 are ignored even if you filled them in, which is a nice way to build out a maxed-out stack and then dial slots on and off without deleting nodes.

The one input you can't skip is model - it has to come from Nunchaku FLUX DiT Loader specifically. This node doesn't load or dequantize anything itself; it patches whatever quantized model you hand it. Nunchaku here means the SVDQuant 4-bit engine, which is real and widely used precisely because it makes FLUX (and Qwen-Image) fast on consumer GPUs without the usual quality hit of naive 4-bit quantization - people report it running comparably to fp8 GGUF but noticeably quicker.

For each slot you get lora_name_X and then a choice of strength controls depending on input_mode. In simple mode it's one lora_wt_X per slot - the overall strength. In advanced mode that splits into model_str_X and clip_str_X so you can weight the diffusion-model patch and the CLIP patch separately. Worth knowing: this node only outputs a modified MODEL, no CLIP output at all, even in advanced mode - which tracks with how most FLUX LoRAs are actually trained, targeting the transformer blocks rather than the text encoder, so the CLIP-strength knob here is really for the rare LoRA that does touch text conditioning. All the strength fields run from -100 to 100 and default to 1, and yes, negative values are legal if you want to subtract a LoRA's effect rather than add it.

Wire the single MODEL output straight into your KSampler chain (or into the next LoRA-stacking node if you're chaining packs).

Installing it: through ComfyUI Manager, search "ComfyUI-NunchakuFluxLoraStacker," or clone it by hand:

cd ComfyUI/custom_nodes
git clone https://github.com/ussoewwin/ComfyUI-NunchakuFluxLoraStacker.git

Restart ComfyUI after. The one dependency that actually matters here is the separate nunchaku core package - install that in your Python environment before this pack does anything useful, since this node is entirely downstream of a Nunchaku-loaded model.

Where people get stuck: the single most common failure is not this node at all - it's forgetting the model has to originate from Nunchaku's own DiT loader, not a stock checkpoint loader. Feed it a regular FLUX model and it'll either error or silently do nothing useful, because the LoRA patching logic here assumes a quantized state dict. Beyond that, the usual FLUX-LoRA gotchas apply: a LoRA trained for one FLUX variant (dev vs schnell, or a distilled fork) can behave oddly on another, and stacking several strong LoRAs at once is unpredictable enough that it's worth testing one at a time with a fixed seed before you blame the stack as a whole.