Bernini-R Load LoRA

Bernini-R is fine-tuned from Wan weights, so the obvious question - "do my Wan LoRAs work?" - has a genuinely messy answer, and BerniniR_LoadLoRA sits right in the middle of it. This node queues a LoRA onto your Bernini-R model handle the way the rest of this pack works: nothing is merged at the node, the spec is appended to the handle and applied when the sampler finally loads the weights.

How it works

In the old world, a LoRA loader loads the model, merges the LoRA, and hands you a patched model. Here, the loader returns a lightweight BERNINI_MODEL_HANDLE, this node appends a LoRA spec (file + strength) to it, and the merge happens lazily at sampling time alongside the model load. You can chain multiple LoRA nodes in sequence - each one just appends another spec - and the sampler applies them in order. The node's description says it handles ComfyUI, Diffusers, Kohya, LyCORIS, and Fun LoRA key formats, which is a nice breadth given how the Wan LoRA ecosystem is a pile of converters.

The inputs

• model_handle - from BerniniR_ModelLoader (or the output of a previous LoadLoRA, if you're stacking).
• lora_name - the LoRA file from ComfyUI/models/loras/.
• strength_model - LoRA strength for the diffusion model, default 1.0. Can go negative (up to ±100), which is a real technique for subtraction-style LoRAs, though ±1 to ±2 is the sane range for almost everything.

Output and wiring

One output: the model_handle again, now carrying the LoRA spec. Pass it through BerniniR_CompileModel if you want, then into the sampler. Order of compile vs. LoRA genuinely doesn't matter here because both just stamp config on the handle.

The honest caveat about Wan LoRAs on Bernini

This is the part you should read twice. Because Bernini-R shares the Wan 2.2 backbone, speed-distillation LoRAs (lightx2v, CausVid) and Wan 2.2 LoRAs partially transfer - Kijai, who wrote the mainline ComfyUI Bernini integration, confirmed lightx2v works "to some extent," and that's the community's most reliable datapoint. But "to some extent" is doing a lot of work. People report likeness actually getting worse than plain Wan 2.2 once LoRAs are stacked, and distilled LoRAs are a known quality killer on the base model (the "speed LoRAs destroy everything that's good in it" thread was about exactly this). Start at low strength, and don't expect character LoRAs trained on Wan 2.2 to hit the same identity out of the box. The model was fine-tuned away from the vanilla Wan weights, and LoRA layers tuned for the base don't always map cleanly onto the fine-tune.

Installation context

The pack installs via ComfyUI Manager (search "ComfyUI-BerniniR Wrapper") or cd ComfyUI/custom_nodes && git clone https://github.com/xiaolibai-sys/ComfyUI-BerniniRWrapper, then restart. Dependencies are minimal (numpy, einops, tqdm, comfy-kitchen, plus pywin32 on Windows) - LoRAs don't add any. Just drop your .safetensors LoRAs in loras/ and they'll show up in the dropdown.

Where people get burned: the dropdown not refreshing after you add a LoRA (restart ComfyUI or refresh node definitions), and cranking strength_model thinking more is more. On a derivative model like this, subtle beats aggressive.