Skip to content

Load LoRA

Loads a LoRA file from local disk and exposes it as an output that the Pipeline Builder accepts.

Category: ModularDiffusion/Pipeline

TL;DR

  • One LoRA per node. To stack LoRAs, drop multiple Load LoRA nodes and connect them all to the same consumer.
  • Two ways to use a LoRA:
    • Pipeline Builder — LoRA is fused (baked) into the model weights at build time. weight is fixed at that point; changing it rebuilds the cached pipeline. Best for LoRAs you always want active.
    • LoRA Pipeline — LoRA is activated per generation and released afterward. The base pipeline is never modified, so changing weight between runs does not trigger a rebuild. Best for IC LoRAs, slider LoRAs, or workflows that swap adapters between branches.
  • Accepts .safetensors, .sft, .pt, .bin, .json, .lora.

Typical workflow position

# Fused (baked into model at build time):
[Load LoRA] ─┐
[Load LoRA] ─┼─→ Pipeline Builder → Generate Media Latents
[Load LoRA] ─┘

# Per-generation (applied at run time):
Pipeline Builder ──┐
[Load LoRA] ───────┤→ LoRA Pipeline → Generate Media Latents
[Load LoRA] ───────┘

Node preview

Load LoRA

Inputs

Name Type Required Notes
file_path path Yes Absolute path to the LoRA file.
weight float (0.0–1.0) No Influence of this LoRA, default 1.0.

Outputs

Name Type Notes
loras loras (dict) {path: weight} — connect to the loras input on the Pipeline Builder.
trigger_phrase str Optional pass-through phrase to include in your prompt; hidden by default.

Tips & pitfalls

  • Hugging Face repo IDs are not supported here. Download the file first; this node loads from disk only.
  • LoRA must match the base pipeline architecture (Flux LoRA → Flux pipeline, etc.). Mismatches surface at pipeline-build time, not when the LoRA loads.
  • weight is baked in at fuse time. Because LoRAs are fused into the model, you cannot change weight between generations without triggering a full pipeline rebuild.
  • Trigger phrases: if your LoRA needs a trigger word, put it in your prompt manually — the trigger_phrase parameter is hidden by default and is currently passthrough metadata only.

See also