InstaSD API Input - Lora Loader

Let's be honest about what InstaCLoraLoader is: a reimplementation of ComfyUI's built-in LoraLoader. Same inputs, same outputs, same comfy.sd.load_lora_for_models call under the hood. It adds descriptive tooltips and a wider strength range, but it does not add features. The only real reason to use it is consistency - if you're building a workflow to publish through InstaSD, this is the LoRA loader the pack's templates expect, so your published workflow doesn't pull in node packs the runner may not have.

If you're not heading to InstaSD, the built-in LoraLoader is a perfectly good choice and you can stop reading. Still here? Fine - here's how it works.

What you set

• model - the diffusion model, from a checkpoint or model loader.
• clip - the CLIP model that goes with it.
• lora_name - a dropdown of every .safetensors in your models/loras folder. This is the whole point of a loader: pick the file, it's applied.
• strength_model - how hard the LoRA modifies the diffusion model, default 1.0, range -100 to 100, step 0.01. Negative values actively un-apply the LoRA's influence.
• strength_clip - same knob for the CLIP/text side, default 1.0.

Outputs are MODEL and CLIP, wired forward to your sampler - the modified model, and the modified CLIP. If both strengths are 0 it returns the inputs untouched (a no-op, but it saves the load). LoRAs are patches, not full models: they adjust cross-attention weights on top of a checkpoint, which is why a 10–200 MB file can restyle an entire generation.

Two community habits worth copying: LoRA strength of 1.0 is often too strong - 0.5–0.8 is where most people settle, and negative strengths are a legit tool for suppressing a baked-in style. And you can chain loaders: stack a character LoRA on top of a style LoRA by feeding one loader's output into the next. That stacking is explicitly what the author's description calls out ("Multiple LoRA nodes can be linked together").

Install

cd ComfyUI/custom_nodes
git clone https://github.com/WaddingtonHoldings/ComfyUI-InstaSD

Restart ComfyUI, or use ComfyUI Manager (search "InstaSD"). Dependencies are boto3, Pillow, requests, numpy. No LoRA files ship with the pack - you supply those from CivitAI or wherever you train.

Gotchas

• It's a clone. If the built-in loader gives you an error, this one will too - same code path. Don't switch to it expecting different behavior.
• The lora dropdown reads your local models/loras. In an InstaSD deployment, the runner needs the LoRA file on its machine or the workflow fails on load.
• Strength range is ±100, which is generous to the point of absurd - going above ~1.5 on most LoRAs just produces mush. The range is a leash, not a target.