Easy Apply PuLID(applyPulID)

Easy Apply PuLID applies a PuLID identity adapter to the model using a reference face image and optional attention mask. It is designed as a simple, single-node interface for the adapter workflow: supply your base model, a PuLID model file, a face image, and control how strongly the identity is injected.

Inputs

• model (MODEL): The base diffusion model to which the PuLID adapter will be applied.
• pulid_file (COMBO): The PuLID model file to load. Select one of the available .safetensors/model files bundled or configured in the node.
• insightface (COMBO): The execution device used by InsightFace for face analysis and embedding extraction. Options are CPU, CUDA, and ROCM. Choose CUDA or ROCM for GPU acceleration, or CPU if no compatible GPU is available.
• image (IMAGE): Reference image containing the face whose identity should be transferred. The face is detected and encoded using InsightFace.
• method (COMBO): Injection mode for the adapter. Options:
  • fidelity: Prioritize strong identity preservation.
  • style: Prioritize the style of the generated output.
  • neutral: Use a balanced middle ground between fidelity and style.
• weight (FLOAT): Overall strength of the PuLID influence. Default is 1. Range is -1 to 5, adjustable in steps of 0.05. Values above 1 intensify identity features; values at or below 0 reduce or invert the adapter's effect.
• start_at (FLOAT): Timestep fraction at which the adapter begins applying its influence. Default is 0 (start of sampling).
• end_at (FLOAT): Timestep fraction at which the adapter stops applying its influence. Default is 1 (end of sampling).
• attn_mask (MASK, optional): An attention mask that limits where the PuLID adapter affects the generated image. Useful for applying identity only to specific regions of the latent space.

Outputs

• model (MODEL): The modified model with the PuLID adapter applied, ready to be connected to a sampler or further adapter nodes.

Usage Notes

• The node expects a clear, forward-facing face in the reference image for best embedding quality.
• Use start_at and end_at to scope the adapter's effect to a portion of sampling, which can help preserve composition early on while enforcing identity later.
• The attn_mask input is optional. When omitted, the adapter affects the whole image.
• If InsightFace cannot detect a face in the reference image, the node will fail or produce an invalid embedding, so crop or align the face beforehand if needed.