Nunchaku FLUX LoRA Loader (URL/Path)

Nunchaku is MIT's SVDQuant quantization for FLUX: it packs the diffusion model down to roughly 4-bit precision and, on the right hardware, makes FLUX generation dramatically faster with barely a quality hit. It's community-driven rather than official, and it's real - people running Nunchaku-quantized Chroma with PuLID have reported full-image generations in around ten seconds. The catch with LoRAs on a quantized model is that a standard LoRA file often isn't directly usable; it needs Nunchaku-specific conversion before it can patch a 4-bit model correctly. This node's own description spells out exactly that: it "loads LoRAs from URLs or local paths with caching and applies Nunchaku-specific conversion." That conversion step, done automatically at load time instead of as a separate offline script, is the entire reason this node exists rather than you just using the regular LoRA loader from elsewhere in the same pack.

What it needs, and what it gives back

• model (required) - the diffusion model to patch. The tooltip is explicit and worth repeating verbatim: "Make sure the model is loaded by Nunchaku FLUX DiT Loader." That loader node is not part of this pack - it comes from the separate ComfyUI-nunchaku pack (mit-han-lab). If you don't already have that installed and a Nunchaku FLUX DiT Loader upstream feeding a quantized model into this node, it has nothing correctly-formatted to patch.
• lora_url - the URL or local path to the LoRA file. Same fetch-at-run-time behavior as everything else in this pack: paste a link, skip the manual download.
• lora_strength - how strongly the LoRA modifies the model, and it can go negative. The slider technically allows anywhere from -100 to 100, which is far wider than the usual LoRA range - don't read that as guidance. Treat it the way you'd treat any LoRA strength: start around 1.0 and adjust from there. Community weight conventions for LoRAs generally sit in the 0.5–1.5 range depending on the base model and the specific LoRA's own recommendation, and there's no reason a Nunchaku-converted one changes that.

The output is a single MODEL - the modified diffusion model, ready to feed into your sampler exactly like any patched model would be.

Installing it

Search ComfyUI-load-lora-from-url in ComfyUI Manager, or:

cd ComfyUI/custom_nodes
git clone https://github.com/bollerdominik/ComfyUI-load-lora-from-url

This node is useless on its own - you also need ComfyUI-nunchaku installed and a FLUX checkpoint already loaded through its Nunchaku FLUX DiT Loader, plus (per Nunchaku's own setup) the matching nunchaku Python wheel for your CUDA/torch version. None of that ships with this small pack; it's a separate stack you build first.

Where people get burned

• Feeding it a model that wasn't loaded through Nunchaku. The tooltip's warning is the whole failure mode: if model didn't come from the Nunchaku FLUX DiT Loader, this node either errors out or produces garbage, because it's expecting the quantized representation, not a normal fp16/fp8 checkpoint.
• A LoRA that isn't compatible with the base FLUX architecture. Quantization doesn't change the underlying rule - a LoRA trained for a different base model still won't transfer cleanly onto FLUX just because this node can technically load the file.
• Nunchaku's own install friction. The wheel is version-pinned to specific torch/CUDA combinations, and mismatches there are the most common source of import errors with anything Nunchaku-related, independent of this node entirely. If FLUX loads fine through the stock loader but this whole chain errors on startup, check the Nunchaku pack's own install first before assuming this node is at fault.