FLUX LoRA Loader V2

This is the node the author actually wants you using now. If you looked at FluxLoraMultiLoader first and noticed ComfyUI labels it "Legacy - Do Not Use in V2," this - "FLUX LoRA Loader V2" - is where that road leads. It was built specifically because ComfyUI's newer Nodes 2.0 / Desktop frontend changed how widget management works under the hood, and the old node's dynamic-slot trick didn't survive the transition cleanly.

Functionally it does the same job as the legacy node: apply up to ten LoRAs to a Nunchaku FLUX model in a single node. The interface got simpler in the rewrite, though. There's no input_mode toggle and no separate model/CLIP strength split anymore - every slot is just lora_name_X plus one lora_wt_X, both optional, and both default to doing nothing (None filename, weight 1.0 if you do pick a file). There's also no backend lora_count widget in this version; slot visibility is a frontend-only combo box - "🔢 LoRA Count" - and the nodes you don't show simply don't get sent, so an empty slot costs you nothing and throws no validation error. That's the actual improvement V2 delivers over V1: no more required-but-unused fields cluttering the graph, and the node resizes itself to fit only the slots you're using.

The only required input is model, and same rule as always with this pack: it has to be the output of Nunchaku FLUX DiT Loader, not a stock checkpoint loader. Nunchaku here is the SVDQuant 4-bit quantization engine - real, actively maintained tech (nunchaku.tech) that's become one of the standard ways people run FLUX and Qwen-Image fast on consumer cards without the quality collapse you'd expect from naive 4-bit. If your model input isn't coming from that loader, this node has nothing valid to patch.

Strength range is generous: -100 to 100 per slot, step 0.001 (tighter than V1's 0.01, if you're the type who fine-tunes to three decimal places). Output is a single MODEL - no CLIP, because FLUX LoRAs overwhelmingly target the transformer/DiT blocks rather than the text encoder, so there's nothing meaningful to patch there in the common case. Wire that MODEL straight into your sampler or the next node in the chain.

Install: search "ComfyUI-NunchakuFluxLoraStacker" in ComfyUI Manager, or clone it directly:

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

Restart after. You also need the nunchaku Python package installed separately - this pack only patches models Nunchaku already loaded, it doesn't bundle the quantization engine itself.

Troubleshooting: the recurring one across the whole Nunchaku-FLUX ecosystem, not just this node, is a black or blank output when sage_attention is enabled at ComfyUI launch - people chasing this on Nunchaku's Qwen-Image loaders have consistently traced it to --use-sage-attention (or the equivalent launch flag) fighting with Nunchaku's own attention path, and the fix is to drop it from your startup args rather than disable it per-node. If your LoRA seems to do nothing at all, double check the filename actually resolved (an empty dropdown usually means the file isn't in your models/loras folder, or isn't compatible with the FLUX architecture this node expects) before assuming the weight is wrong.