Blender Input Load LoRA

LoRAs are how you bend a model to a style or a subject, and when you're working in Blender you want to switch and mix them without leaving the viewport. BlenderInputLoadLora puts the LoRA picker on the add-on's panel: a dropdown of every LoRA on the ComfyUI server, plus a strength slider. And because it outputs a MODEL you can feed into another BlenderInputLoadLora, you can stack several - the pack's own description calls this out explicitly, and it's the workflow pattern you'll want for "style LoRA + character LoRA."

The source is a subclass of ComfyUI's LoraLoaderModelOnly, which is the model-only variant - it takes an already-loaded MODEL and applies the LoRA to it, without touching CLIP or VAE. That matters: unlike the fuller LoraLoader, this node won't also remap your text encoder, which is fine for style/character LoRAs that only affect denoising and keeps the graph simpler. Its execute normalizes the path and calls the parent's load_lora_model_only.

Inputs and outputs

• model - the MODEL input to apply the LoRA to. Feed it your checkpoint or diffusion-model output, or the output of another BlenderInputLoadLora in the chain.
• lora_name - which LoRA, from the server's loras folder. A dropdown in Blender.
• strength_model - how hard the LoRA bites, 0 to 1 (the range goes to ±100 if you want to overdrive or invert). Start at 1.0, back it off when things get crunchy.
• default - which LoRA the dropdown starts on; empty means none selected.
• order - panel position, lower first.

Output is a single MODEL - the modified diffusion model, fed to the next loader in the chain or into the sampler. Note there's no CLIP output, unlike the full LoraLoader; the pack made the model-only choice deliberately.

Installing it

Standard for the pack. ComfyUI side, ComfyUI Manager → search "ComfyUI-Blender", or:

cd ComfyUI/custom_nodes
git clone https://github.com/alexisrolland/ComfyUI-Blender

Restart ComfyUI. No extra Python dependencies, but you need a current ComfyUI - schema v3 nodes won't load on older builds. The Blender add-on is the other half, a ZIP from the latest release installed via Edit > Preferences > Add-ons > Install from Disk.

Where people get burned

The dropdown lists the server's loras folder, so a remote ComfyUI means remote LoRAs - put the files on the box that runs ComfyUI. Second, if your workflow relies on a LoRA that changes prompt behavior (one with trigger-word-heavy training), the model-only loader won't apply any CLIP remap, so prompt-side effects won't come along; that's expected, not a bug. Third, chain order matters: the LoRA closest to the checkpoint applies first, and strengths compound, so a second LoRA at 1.0 layered on an already-strong first one can oversaturate fast - back them both off. Panel only builds after an API-format export and import, and the title is the label ("Style LoRA" beats "lora3"). Same-GPU VRAM: each stacked LoRA adds a little, and the viewport already has a claim on the card.