FaceSIM weights (unpublished; do not redistribute)
Expression-similarity encoders fine-tuned on Ava-256 avatar-capture correspondences (synchronized views, and
cross-identity neighbours from the released avatar expression codes). Code: GitHub seanmamasde/facesim (private).
The metric is the cosine distance between backbone CLS embeddings of two images,
d(x, y) = 1 - cos(f(x), f(y)), on full frames resized (squashed) to 224×224, inputs in [0, 1]
(the backbone wrapper applies its own normalization; the ImageNet ViT deliberately uses none).
| folder | model | seeds | FE70k SRCC (all / held-out) | FEC acc. |
|---|---|---|---|---|
imagenet_vitb16/ours_s{1,2,3} |
ImageNet ViT-B/16, synchronized views + code neighbours (main model) | 3 | 0.865 / 0.856 | 44.3 |
imagenet_vitb16/sync_only_s{1,2,3} |
ImageNet ViT-B/16, synchronized views only | 3 | 0.842 / 0.829 | 43.7 |
dinov2_vitb14/sync_only_s{1,2,3} |
DINOv2 ViT-B/14, synchronized views only, layer-wise lr decay | 3 | 0.805 / 0.805 | 57.8 |
dinov2_vitl14/sync_only_s{1,2} |
DINOv2 ViT-L/14, synchronized views only | 2 | 0.822 / 0.824 | 61.0 |
imagenet_vitb16/fec_supervised_reference_s{1,2,3} |
FEC-supervised ViT reference (human FEC triplets, FECNet triplet loss, margins 0.1 / 0.2) | 3 | 0.793 / 0.788 | 78.6 |
Each folder: final.pth (state dict of the student: backbone.*, projector.*, predictor.*; for the FEC reference
only backbone.*), args.json (training arguments), metrics.json (FE70k all / dev / held-out actors, FEC test).
All checkpoints are the last training step (no checkpoint selection), except the FEC reference, which is selected on
FEC development triplets.
# with the GitHub code on PYTHONPATH
from eval_ckpt import load_backbone
f = load_backbone("imagenet_vitb16/ours_s1/final.pth").cuda().eval() # reads args.json next to the file
z = f(images) # images: float [B, 3, 224, 224] in [0, 1]