Identity Preservation

Identity preservation is a model's ability to keep a specific subject recognizable as the same individual across its outputs: the same face across a set of generated portraits, the same character from the first frame of a generated video to the last, or the same tracking ID for a pedestrian who walks behind a bus and reappears. Two communities use the term. In generative AI, it describes personalization methods that place a real person or product into new scenes without changing who or what it is. In multi-object tracking (MOT), it describes keeping each object's ID stable through occlusions, crossings, and camera changes. Both measure it with identity embeddings, feature vectors that land close together for the same individual and far apart for different ones.

Identity Preservation in Image and Video Generation#

A text-to-image model knows what "a woman" or "a dog" looks like but not what one particular woman or dog looks like, so personalization methods add that knowledge. Fine-tuning methods learn it from a few photos. Google's DreamBooth (2022) fine-tunes the whole diffusion model on three to five images of a subject tied to a rare token, with a class-specific prior preservation loss that keeps the model generating other dogs normally, and LoRA adapters make the same idea cheap enough to train per person.

Encoder-based methods skip per-subject training. Tencent's IP-Adapter (2023) feeds image embeddings into the diffusion model through a separate cross-attention layer, and InstantID (January 2024) preserves a face from a single photo by combining an embedding from a face recognition network with IdentityNet, a ControlNet conditioned on five facial landmarks. Stronger identity conditioning makes the output look more like the reference and also copies more of its pose, expression, and lighting, so these tools expose a strength setting that trades likeness against how far the prompt can change the scene.

Generated faces are scored by the cosine similarity between ArcFace embeddings of the reference photo and the output; ArcFace, introduced by Jiankang Deng and colleagues at CVPR 2019, trains face embeddings with an angular margin so that identities separate cleanly on a hypersphere. Video adds a second requirement, keeping the subject consistent across frames, which VBench measures as subject consistency by comparing DINO features between frames. In video generation, a character whose face drifts over a few seconds fails this test even when every single frame looks realistic.

Identity Preservation in Tracking#

A tracker assigns IDs by matching each frame's detections to existing tracks. Motion alone, a Kalman filter predicting where each box will be plus an intersection over union (IoU) match, works until two people cross or one disappears behind an obstacle; then the tracker swaps their IDs or starts a new one, an error counted as an identity switch. Appearance features fix much of this. Nicolai Wojke, Alex Bewley, and Dietrich Paulus's DeepSORT (2017) added embeddings from a network trained on a person re-identification dataset to the motion-only SORT tracker and cut identity switches by 45 percent. The IDF1 metric, proposed by Ergys Ristani and colleagues in 2016, scores identity preservation directly as the share of detections assigned their correct identity.

Ultralytics YOLO tracking defaults to TrackTrack (Shim and colleagues, CVPR 2025), which keeps lost tracks alive for 30 frames (track_buffer) so a returning object can reclaim its ID and suppresses duplicate new IDs before creating one. Setting with_reid: True in the tracker YAML adds appearance matching, by default with the detector's own features and optionally with a dedicated encoder such as the 2.8-million-parameter YOLO26n-reid model. The snippet tracks people in a sample video with and without ReID and counts the distinct IDs each run assigns:

from ultralytics import YOLO
from ultralytics.utils import ROOT, YAML

# Copy the default TrackTrack config and turn on appearance re-identification
cfg = YAML.load(ROOT / "cfg/trackers/tracktrack.yaml")
cfg.update(with_reid=True, model="yolo26n-reid.onnx")
YAML.save("tracktrack_reid.yaml", cfg)

video = "https://github.com/ultralytics/assets/releases/download/v0.0.0/decelera_portrait_min.mov"
for tracker in ["tracktrack.yaml", "tracktrack_reid.yaml"]:
    model = YOLO("yolo26n.pt")  # fresh model so each run starts with no tracks
    ids = set()
    for result in model.track(video, tracker=tracker, classes=[0], stream=True, verbose=False):
        if result.boxes.is_track:
            ids.update(result.boxes.id.int().tolist())
    print(f"{tracker}: {len(ids)} person IDs")

On this short clip both runs assign 3 person IDs, so appearance matching changes nothing here; it matters on longer footage where people leave the frame and return. For the same people, fewer distinct IDs means fewer identity switches, which is what downstream counts depend on: a people counting system that gives a returning shopper a new ID counts that shopper twice.

Related Terms#

• Facial recognition: Names a person by matching their face embedding against a gallery of known people, using the same kind of embedding that identity-preserving generators take as conditioning.
• Data privacy: Redaction and anonymization deliberately remove identity, the opposite goal.
• Deepfakes: Face swaps use identity-preserving generation without the subject's consent.