Generative Vision Atlas

Conditioning & control option

Identity-preserving face conditioning

Combine a semantic face-embedding branch (from a face-recognition encoder) with a spatial landmark/structure branch, injected through some mix of decoupled cross-attention and adapter-style side networks, to preserve a specific person's identity across generations without per-identity fine-tuning.

Mechanism

InstantID preserves a specific person’s identity from a single reference photo, with no per-identity fine-tuning, by “imposing strong semantic and weak spatial conditions” through a dedicated IdentityNet: a face-recognition embedding supplies the semantic signal, and facial landmark images supply a weaker spatial signal, both integrated alongside the text prompt to steer generation on top of an existing frozen checkpoint (SD1.5, SDXL). PuLID takes a different route to the same goal, pairing a standard diffusion branch with a parallel “Lightning T2I” branch and adding a contrastive alignment loss plus an accurate-ID loss, explicitly to minimize how much identity injection disturbs the base model’s other behavior (background, lighting, composition, style).

Trade-offs

Parameter cost: both are adapter-scale additions to a frozen base model, not full fine-tunes — this is what makes “zero-shot… in seconds” (InstantID) and “tuning-free” (PuLID) possible from one photo. Flexibility: both plug into existing community checkpoints rather than requiring a dedicated base model. Precision vs. base-model disruption: the core tension both papers are solving is that a strong enough identity signal tends to leak into and override other aspects of the generation (style, pose, background) unless it is explicitly isolated — InstantID by separating semantic from spatial conditioning strength, PuLID by adding losses specifically penalizing that disruption.

Introduced by

Used by (2)

InstantID, PuLID

Alternatives on this axis

adaLN / adaLN-Zero modulation, Adapter-based spatial conditioning, CLIP text encoder as conditioning source, Cross-attention conditioning, Decoupled cross-attention (image-prompt adapter), Joint (shared) attention conditioning, Native autoregressive token conditioning, Learnable query-token bridge, In-context sequence concatenation, Frozen large-language-model text encoder (T5-style), Vision-language model as text encoder

← All Conditioning & control options