You should look at how Arcface embeddings work. islands on the surface manifold of a hypersphere, everything is unit length from the center
October 25, 2026 at 7:02 AM
You should look at how Arcface embeddings work. islands on the surface manifold of a hypersphere, everything is unit length from the center
📣 New Podcast! "How to Build a Facial Database for OSINT Investigations" on @Spreaker #ai #arcface #artificialintelligence #cyberintelligence #cybersecurity #dataprovenance #deepface #digitalinvestigation #ethicalosint #facenet #facialdatabase #facialembeddings #facialrecognition #faiss #osint
How to Build a Facial Database for OSINT Investigations
https://www.osintinvestigate.com
How can investigators organize public facial images for OSINT research? This episode explores facial-image databases, data provenance, image cleaning, AI embeddings, vector search, verification, privacy, and responsible intelligence practices. Learn why facial similarity should be treated as an investigative lead—not proof of identity.
www.spreaker.com
September 28, 2026 at 3:15 PM
📣 New Podcast! "How to Build a Facial Database for OSINT Investigations" on @Spreaker #ai #arcface #artificialintelligence #cyberintelligence #cybersecurity #dataprovenance #deepface #digitalinvestigation #ethicalosint #facenet #facialdatabase #facialembeddings #facialrecognition #faiss #osint
昔の実装掘り起こして、ArcFace使ってたつもりだったけど今見たらL2-constrained Softmaxだった
September 11, 2026 at 4:38 AM
昔の実装掘り起こして、ArcFace使ってたつもりだったけど今見たらL2-constrained Softmaxだった
Choosing an ArcFace Verifier: Thresholds 0.30 and 0.35 — https://kahma.io/blog/choosing-an-arcface-verifier-thresholds-030-and-035.php
September 2, 2026 at 11:01 AM
Choosing an ArcFace Verifier: Thresholds 0.30 and 0.35 — https://kahma.io/blog/choosing-an-arcface-verifier-thresholds-030-and-035.php
Yes, detection is Vision framework — same stack Photos and Quick Look use. Recognition is ArcFace via Core ML, embeddings clustered locally. Semantic search is MobileCLIP. All three models bundled, all inference on-device, computeUnits = .all so it rides the Neural Engine where available.
June 28, 2026 at 12:28 AM
Yes, detection is Vision framework — same stack Photos and Quick Look use. Recognition is ArcFace via Core ML, embeddings clustered locally. Semantic search is MobileCLIP. All three models bundled, all inference on-device, computeUnits = .all so it rides the Neural Engine where available.
ArcFaceみたいなロスに対応して欲しいと思ったけど、もしかして出来るのかな。見てみよう。
September 12, 2023 at 7:16 AM
ArcFaceみたいなロスに対応して欲しいと思ったけど、もしかして出来るのかな。見てみよう。
Teaching softmax some new tricks with ArcFace’s additive angular margin loss: www.johntrimble.com/posts/from-s...
#MachineLearning #ArcFace
#MachineLearning #ArcFace
From Softmax to ArcFace: Building Better Embeddings with Additive Angular Margins
Introduction
www.johntrimble.com
August 14, 2025 at 5:20 PM
Teaching softmax some new tricks with ArcFace’s additive angular margin loss: www.johntrimble.com/posts/from-s...
#MachineLearning #ArcFace
#MachineLearning #ArcFace
What my lab found really challenging about doing this is the “system” part. The models themselves (esp ArcFace, MagFace, SphereFace etc) have certainly undergone architectural improvements over time and in lab settings there is some data to back up that improvement.
November 23, 2025 at 9:53 PM
What my lab found really challenging about doing this is the “system” part. The models themselves (esp ArcFace, MagFace, SphereFace etc) have certainly undergone architectural improvements over time and in lab settings there is some data to back up that improvement.
✅ Contrastive vs. Classification Losses? If using small batches (less than 256), classification/stateful losses (e.g., ArcFace) work best. If using larger batches (more than 256), contrastive/stateless losses (e.g., NT-Xent/InfoNCE, MultiSimilarity) give better results, but ...
March 18, 2025 at 10:41 PM
✅ Contrastive vs. Classification Losses? If using small batches (less than 256), classification/stateful losses (e.g., ArcFace) work best. If using larger batches (more than 256), contrastive/stateless losses (e.g., NT-Xent/InfoNCE, MultiSimilarity) give better results, but ...
training a ResNet-50 facial feature extraction model based on ArcFace loss on a raw and randomly defrontalized large-scale dataset, where defrontalization was performed with our previously trained face defrontalization model. Our method was compared [5/8 of https://arxiv.org/abs/2506.04496v1]
June 6, 2025 at 5:58 AM
training a ResNet-50 facial feature extraction model based on ArcFace loss on a raw and randomly defrontalized large-scale dataset, where defrontalization was performed with our previously trained face defrontalization model. Our method was compared [5/8 of https://arxiv.org/abs/2506.04496v1]
FairFace, we simulate degraded images and evaluate performance using Deepface with ArcFace loss in 1:n identification tasks. We perform an experiment and find that false positive rates peak near baseline image quality, while false negatives increase [3/7 of https://arxiv.org/abs/2505.14320v1]
May 21, 2025 at 6:06 AM
FairFace, we simulate degraded images and evaluate performance using Deepface with ArcFace loss in 1:n identification tasks. We perform an experiment and find that false positive rates peak near baseline image quality, while false negatives increase [3/7 of https://arxiv.org/abs/2505.14320v1]
(LaPIG). This framework enables the construction of comprehensive, high-quality paired visible and thermal images using captions generated by LLMs. Our method encompasses three parts: visible image synthesis with ArcFace embedding, thermal image [3/5 of https://arxiv.org/abs/2503.16376v1]
March 21, 2025 at 6:08 AM
(LaPIG). This framework enables the construction of comprehensive, high-quality paired visible and thermal images using captions generated by LLMs. Our method encompasses three parts: visible image synthesis with ArcFace embedding, thermal image [3/5 of https://arxiv.org/abs/2503.16376v1]
All You Need to Know About Training Image Retrieval Models
@berton-gabri.bsky.social , Kevin Musgrave, Carlo Masone
tl;dr: use ArcFace with small batch size or hard contrastive/triplet with huge batch size. Noisy labels no problem, bigger size is better
arxiv.org/abs/2503.13045
@berton-gabri.bsky.social , Kevin Musgrave, Carlo Masone
tl;dr: use ArcFace with small batch size or hard contrastive/triplet with huge batch size. Noisy labels no problem, bigger size is better
arxiv.org/abs/2503.13045
March 19, 2025 at 9:45 AM
All You Need to Know About Training Image Retrieval Models
@berton-gabri.bsky.social , Kevin Musgrave, Carlo Masone
tl;dr: use ArcFace with small batch size or hard contrastive/triplet with huge batch size. Noisy labels no problem, bigger size is better
arxiv.org/abs/2503.13045
@berton-gabri.bsky.social , Kevin Musgrave, Carlo Masone
tl;dr: use ArcFace with small batch size or hard contrastive/triplet with huge batch size. Noisy labels no problem, bigger size is better
arxiv.org/abs/2503.13045
Right, exactly - same conclusion I had with patch descriptors - with 1024+ batch size in 2020, the hard-triplet loss is way better than ArcFace. Good to have a paper reference for than now
x.com/ducha_aiki/s...
x.com/ducha_aiki/s...
March 19, 2025 at 9:13 AM
Right, exactly - same conclusion I had with patch descriptors - with 1024+ batch size in 2020, the hard-triplet loss is way better than ArcFace. Good to have a paper reference for than now
x.com/ducha_aiki/s...
x.com/ducha_aiki/s...
Facial Recognition in C# with ONNX and ArcFace: Theory, Math, and Implementation https://old.reddit.com/r/dotnet/comments/1ujoj9b/facial_recognition_in_c_with_onnx_and_arcface/
#dotnet
#dotnet
June 30, 2026 at 1:04 PM
Facial Recognition in C# with ONNX and ArcFace: Theory, Math, and Implementation https://old.reddit.com/r/dotnet/comments/1ujoj9b/facial_recognition_in_c_with_onnx_and_arcface/
#dotnet
#dotnet
DeepFace is a lightweight face recognition and facial attribute analysis (age, gender, emotion and race) framework for python. It is a hybrid face recognition framework wrapping state-of-the-art models: VGG-Face, FaceNet, OpenFace, DeepFace, DeepID, ArcFace, Dlib, SFace and GhostFaceNet.
GitHub - serengil/deepface: A Lightweight Face Recognition and Facial Attribute Analysis (Age, Gender, Emotion and Race) Library for Python
A Lightweight Face Recognition and Facial Attribute Analysis (Age, Gender, Emotion and Race) Library for Python - serengil/deepface
github.com
January 6, 2025 at 2:02 PM
DeepFace is a lightweight face recognition and facial attribute analysis (age, gender, emotion and race) framework for python. It is a hybrid face recognition framework wrapping state-of-the-art models: VGG-Face, FaceNet, OpenFace, DeepFace, DeepID, ArcFace, Dlib, SFace and GhostFaceNet.
Gonna start your new playlist off with this one. See if you can pick where it got sampled.
youtu.be/gGNhgn9fMi8?...
youtu.be/gGNhgn9fMi8?...
A pilgrim (Pielgrzym) - Niemen i Aerolit Sopot 1974
YouTube video by arcface
youtu.be
October 18, 2024 at 5:51 AM
Gonna start your new playlist off with this one. See if you can pick where it got sampled.
youtu.be/gGNhgn9fMi8?...
youtu.be/gGNhgn9fMi8?...
There's more videos linked in the paper, but they're what we've seen from Sora/Runway/Kling/whatever - the main focus for them seems to be the deepfakes (accomplished through ArcFace-guided similarity alignment, which seems violating on multiple levels), video editing/inserts, and SFX/music gen-sync
October 9, 2024 at 12:56 PM
There's more videos linked in the paper, but they're what we've seen from Sora/Runway/Kling/whatever - the main focus for them seems to be the deepfakes (accomplished through ArcFace-guided similarity alignment, which seems violating on multiple levels), video editing/inserts, and SFX/music gen-sync
Been following you since the “Humble Pie” days. 💙
Humble Pie - A song for Jenny 1971
YouTube video by arcface
youtu.be
November 21, 2024 at 10:36 PM
Been following you since the “Humble Pie” days. 💙
That stack is dialed in. computeUnits = .all for ArcFace on the ANE is the right call — consistent latency matters for UX. How do you handle model versioning when you need to update recognition without invalidating stored embeddings? Re-indexing face data on upgrade is the unglamorous part.
June 28, 2026 at 8:02 PM
That stack is dialed in. computeUnits = .all for ArcFace on the ANE is the right call — consistent latency matters for UX. How do you handle model versioning when you need to update recognition without invalidating stored embeddings? Re-indexing face data on upgrade is the unglamorous part.
What software did you use? I tried AWS Rekognition, Facenet512, VGG-Face, and ArcFace. All of them agreed the 2 faces don't match.
April 18, 2025 at 2:33 AM
What software did you use? I tried AWS Rekognition, Facenet512, VGG-Face, and ArcFace. All of them agreed the 2 faces don't match.
A review of Face Recognition loss functions, exploring advancements from ArcFace to modern adaptive and prototype-based methods for improved accuracy. #facerecognition
Why AI Needs to Lose (a Little) to Recognize Your Face Better
hackernoon.com
March 13, 2025 at 12:06 PM
A review of Face Recognition loss functions, exploring advancements from ArcFace to modern adaptive and prototype-based methods for improved accuracy. #facerecognition
Dylan Waldner, Shyamal Mitra
Pairwise Discernment of AffectNet Expressions with ArcFace
https://arxiv.org/abs/2412.01860
Pairwise Discernment of AffectNet Expressions with ArcFace
https://arxiv.org/abs/2412.01860
December 4, 2024 at 1:41 PM
Dylan Waldner, Shyamal Mitra
Pairwise Discernment of AffectNet Expressions with ArcFace
https://arxiv.org/abs/2412.01860
Pairwise Discernment of AffectNet Expressions with ArcFace
https://arxiv.org/abs/2412.01860
CosFace and ArcFace Loss were investigated for text-independent speaker identification using a modified VGG16 model with mel spectrogram inputs; both losses demonstrated superior identification accuracy compared to Softmax loss.
Text-Independent Speaker Identification Using Audio Looping With Margin Based Loss Functions
Elliot Q C Garcia, Nicéias Silva Vilela, Kátia Pires Nascimento do Sacramento, Tiago A. E. Ferreira
arxiv.org
September 30, 2025 at 7:45 AM
CosFace and ArcFace Loss were investigated for text-independent speaker identification using a modified VGG16 model with mel spectrogram inputs; both losses demonstrated superior identification accuracy compared to Softmax loss.
Ascend NPUs now accelerate ArcFace for fp16, low-latency biometrics. However, sparse documentation & specific preprocessing could curb wider adoption. Wider developer uptake feels contingent on a smoother integration path.
High-Performance Biometrics for the Ascend Ecosystem: Discovering the ArcFace Model
aichina.news
May 10, 2026 at 3:44 PM
Ascend NPUs now accelerate ArcFace for fp16, low-latency biometrics. However, sparse documentation & specific preprocessing could curb wider adoption. Wider developer uptake feels contingent on a smoother integration path.