Easy Apply PuLID (Advanced)(applyPulIDADV)

Description

The Easy Apply PuLID (Advanced) node applies PuLID face-identity conditioning to a diffusion model. It is the advanced variant of the standard Easy Apply PuLID node, adding extra controls for identity projection, fidelity, noise, scheduling, and optional attention masking.

Inputs

• model (MODEL, required): The base diffusion model that receives the PuLID identity injection.

• pulid_file (COMBO, required): Selects the PuLID model file used to encode the input identity.

• insightface (COMBO, required): The backend used for face detection and recognition. Options: CPU, CUDA, ROCM.

• image (IMAGE, required): The source image(s) containing the face identity to apply.

• weight (FLOAT, default 1.0, range -1.0 to 5.0): The strength of the identity injection. Higher values strengthen the effect, lower values weaken it, and negative values push away from the provided identity.

• projection (COMBO, default ortho_v2): The projection used on the identity embedding. ortho_v2 is the default, ortho is the previous orthogonal projection, and none disables projection.

• fidelity (INT, default 8, range 0 to 32): Controls how faithfully the injected face preserves the input identity. Higher values generally increase identity resemblance.

• noise (FLOAT, default 0.0, range -1.0 to 1.0): Adds noise to the identity embedding. This can be used to add variation while keeping the same identity.

• start_at (FLOAT, default 0.0, range 0.0 to 1.0): The denoising step fraction at which PuLID application begins.

• end_at (FLOAT, default 1.0, range 0.0 to 1.0): The denoising step fraction at which PuLID application ends.

• attn_mask (MASK, optional): An optional attention mask that restricts the PuLID injection to specific image regions.

Outputs

None — the updated source defines no output sockets. The PuLID application is performed on the supplied model; no separate MODEL output is exposed in this version.

Usage Notes

• PuLID is applied between start_at and end_at of the denoising schedule. With the defaults, it is active for the entire sampling process.
• To inject identity only during the later part of sampling, set start_at above 0.0. To stop applying it before the final steps, set end_at below 1.0.
• The insightface backend should match your runtime: CUDA for NVIDIA GPUs, ROCM for AMD GPUs, or CPU for CPU-only operation.
• Higher fidelity values generally produce stronger identity preservation, while lower values allow more flexibility in expression, pose, and style.
• weight interacts with fidelity: increasing both can produce a stronger identity lock, while reducing one or both gives the model more freedom.
• Using projection = "none" skips the orthogonal projection and may be useful for comparing raw identity embedding behavior.
• When an attn_mask is supplied, the identity injection is localized to the masked region instead of applying globally.