Skip to content

LoRA Pipeline

Layers non-fused LoRA adapters on top of an existing pipeline so the same cached model can power multiple branches simultaneously.

Category: ModularDiffusion/Pipeline

TL;DR

  • Connect a Pipeline Builder output and one or more Load LoRA nodes here; wire the lora_pipeline output to Generate Media Latents instead of the Pipeline Builder directly.
  • Adapters are activated per generation and released afterward — the base pipeline is never permanently modified, so a single cached model can drive multiple branches (one with LoRAs, one without) without rebuilding.
  • Prefer this node over fusing LoRAs on the Pipeline Builder when swapping adapters between runs or using in-context (IC) LoRAs, distillation/acceleration LoRAs, or slider LoRAs.

Typical workflow position

Pipeline Builder ──→ [LoRA Pipeline] ←── Load LoRA
                          │
                          └──→ Generate Media Latents → Decode Media Latent

Node preview

LoRA Pipeline

Inputs

Name Type Required Notes
pipeline Pipeline Config Yes Base diffusion pipeline. Connect from Pipeline Builder. The base pipeline is reused, not modified — wire it to other nodes simultaneously.
loras loras Yes One or more LoRA payloads from Load LoRA nodes. Accepts a list; connect multiple Load LoRA outputs to stack adapters.

Outputs

Name Type Notes
lora_pipeline Pipeline Config Pipeline reference that activates the listed LoRAs around each generation call. Shares the cache entry of the input pipeline — wiring this to Generate Media Latents does not trigger a rebuild.
logs str Build log, including the pipeline configuration hash.

Tips & pitfalls

  • Activation LoRAs vs. fused LoRAs are not equivalent. Fused LoRAs (baked via the Pipeline Builder loras input) are permanently merged into weights; activation LoRAs (this node) are applied transiently. Changing a fused LoRA evicts the entire pipeline cache. Changing an activation LoRA does not.
  • Connect at least one LoRA. The node requires at least one entry in loras to activate — wire one or more Load LoRA nodes before running.
  • The lora_pipeline output shares the cache with the input pipeline. You can wire both to separate Generate Media Latents nodes (one with LoRAs, one without) without loading the model twice.
  • LoRA weights are applied per run. Unlike fused LoRAs, changing weight on a Load LoRA node between runs does not rebuild the pipeline.

See also