☁️BizyAir Load LoRA

The mechanics of a LoRA don't change just because the model lives on someone else's server: it's still a small low-rank patch layered on top of a frozen checkpoint's attention weights, cheap to store, cheap to swap in and out, and, critically, tied to the specific base architecture it was trained against. What this node changes is where that patching happens. Instead of ComfyUI reading a .safetensors file off your disk and merging it into a model already loaded in your VRAM, BizyAir_LoraLoader tells SiliconFlow's cloud "apply this LoRA to that model" and the whole thing, base checkpoint and adapter both, stays server-side.

Inputs that matter

model and clip take BizyAir's own typed sockets (BIZYAIR_MODEL/BIZYAIR_CLIP), so this node only makes sense wired downstream of a BizyAir checkpoint loader, not a stock local one. strength_model and strength_clip do exactly what they do everywhere else in ComfyUI: how strongly the LoRA's patch applies to the diffusion weights versus the text-encoder weights, both default to 1.0, both accept negative values if you want to subtract a LoRA's effect rather than add it.

The two fields worth pausing on are lora_name and model_version_id. lora_name is a dropdown, but unlike a local LoraLoader it isn't reading your models/loras/ folder, it's populated from whatever's registered against your BizyAir account. model_version_id is the more direct route: BizyAir supports uploading and hosting your own custom LoRAs (the README's changelog calls this out explicitly, "share your custom LoRAs" landed back in September 2024), and this field is how you point the loader at a specific version of a model you or someone else has hosted there, rather than picking from a name list.

Outputs are MODEL and CLIP, both in BizyAir's typed flavor, so they chain straight into a BizyAir text encoder and sampler the same way a local LoraLoader's outputs chain into local ones.

The strength question hasn't changed

The old rule of thumb, run LoRAs around 0.5-0.8 strength to avoid overcooking the effect, is specifically SDXL-era advice. On newer flow-matching-family checkpoints the community increasingly runs LoRAs at 1.0 or even above, because dropping strength there tends to cost likeness rather than just softening the effect. Whatever base model this LoRA targets, start from what its own card or usage notes suggest rather than carrying over a number from a different architecture.

Installing it

ComfyUI Manager: search "BizyAir", install, restart. Or clone it directly:

cd ComfyUI/custom_nodes && git clone https://github.com/siliconflow/BizyAir.git

You need a BizyAir/SiliconFlow API key before any of this works, set through the "click to login" prompt on first use. Signup involves phone-number verification, and some users outside China have reported difficulty completing that step, worth planning around before you're mid-workflow.

Common issues

An empty or single-entry lora_name dropdown usually just means your account has nothing registered yet, if you haven't uploaded a custom LoRA or don't have one shared to you, there's nothing for the list to show, that's expected rather than broken. A model_version_id that doesn't resolve will typically surface as a load error at run time rather than failing silently, double-check it against whatever hosting page or upload confirmation gave you that ID.

Mismatched architecture is the classic LoRA failure and it applies here just as much as locally: a LoRA trained for one base model does effectively nothing (or actively degrades output) applied to a BIZYAIR_MODEL from a different architecture, there's no cross-architecture compatibility to expect. And as with every node in this pack, if a run simply doesn't execute, check the API key before assuming the LoRA setup is the problem.

Separately, unrelated to this node's own behavior: ComfyUI Manager occasionally auto-installs the whole BizyAir pack when it misresolves an unrelated missing-node error on an older shared workflow. That's a real, if confusing, quirk rather than anything malicious, and it's removable through the Manager if you didn't intend to install it.