harveymannering/DINOv3_FaceRegionClassifier

DINOv3 Face-Region Classifier

A binary face / non-face classifier finetuned from facebook/dinov3-vits16-pretrain-lvd1689m (DINOv3 ViT-S/16) with a single linear head on the pooled embedding. Given a 224x224 crop, it predicts whether that crop contains a face — including non-human, pareidolia-style "faces" (e.g. two dots and a line in tree bark or rock formations) — making it useful as a guidance signal for steering image generation toward face-like content in a specific image region.

Training data

• ExpWild and iCartoonFace: real human/cartoon face crops
• FacesInThings: pareidolia face crops (faces seen in inanimate objects/textures)
• Negatives: same-size random crops from the same images, sampled to not overlap any labeled face box

See FacePareidolia/classifier for the full training code (train.py, face_dataset.py).

Metrics

Best checkpoint (epoch 4/20): val_loss = 0.061, val_acc = 98.0% (10% image-level holdout).

Usage

The base DINOv3 checkpoint is gated on Hugging Face, so this repo ships its own config.json (architecture metadata only — no gated weights) plus the finetuned weights, letting you rebuild the model without requesting DINOv3 access:

import torch
import torch.nn as nn
from transformers import AutoConfig, AutoModel
from huggingface_hub import hf_hub_download

REPO_ID = "harveymannering/DINOv3_FaceRegionClassifier"

class Dinov3FaceRegionClassifier(nn.Module):
    def __init__(self, config):
        super().__init__()
        self.backbone = AutoModel.from_config(config)
        self.head = nn.Linear(config.hidden_size, 1)

    def forward(self, pixel_values):
        pooled = self.backbone(pixel_values=pixel_values).pooler_output
        return self.head(pooled).squeeze(-1)  # face logit; sigmoid -> p(face)

config = AutoConfig.from_pretrained(REPO_ID)
model = Dinov3FaceRegionClassifier(config)
weights_path = hf_hub_download(repo_id=REPO_ID, filename="dinov3_face_region_classifier.pth")
model.load_state_dict(torch.load(weights_path, map_location="cpu"))
model.eval()

Input should be 224x224 RGB, normalized with ImageNet mean/std (mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]).

License

Derived from facebook/dinov3-vits16-pretrain-lvd1689m; usage is subject to the terms of that model's license — see the base model page for details.