Cosmos-Embed1-448p-anomaly-detection

by nvidia

Video anomaly embeddings for physical-AI monitoring and event retrieval

10Kdl/month

10likes

1.2Bparams

HuggingFace Use in Pipeline

Identifiers

Model ID

nvidia/Cosmos-Embed1-448p-anomaly-detection

Feature URI

mixpeek://video_extractor@v1/nvidia_cosmos_embed1_anomaly_v1

Overview

Cosmos Embed1 Anomaly Detection is a NVIDIA video embedding model tuned for finding unusual events in physical-world video. The HuggingFace model card tags it for video-text retrieval, video embeddings, physical AI, and anomaly detection, making it a strong fit for cameras, robotics, warehouse footage, and field operations where an agent needs to notice abnormal visual behavior.

On Mixpeek, Cosmos Embed1 can turn surveillance or operations video into searchable anomaly features. Agents can retrieve clips that look unsafe, off-process, or visually unusual, then combine those results with object detection, transcript search, or human review workflows.

Architecture

QFormer-based video-text embedder with an EVA-ViT-G visual backbone. The 448p anomaly variant processes 8 sampled video frames into 768-dimensional normalized embeddings, aligns video and text via contrastive training, and is LoRA-fine-tuned on the Vad-Reasoning anomaly dataset.

Mixpeek SDK Integration

import { Mixpeek } from "mixpeek";

const mx = new Mixpeek({ apiKey: "API_KEY" });

await mx.collections.ingest({
  collection_id: "operations-video",
  source: { url: "https://example.com/warehouse-camera.mp4" },
  feature_extractors: [{
    name: "anomaly_detection",
    version: "v1",
    params: {
      model_id: "nvidia/Cosmos-Embed1-448p-anomaly-detection"
    }
  }]
});

Capabilities

Video anomaly scoring for physical-world footage
Video-text retrieval over abnormal events
Fine-tuned Cosmos Embed1 backbone
Works as an agent perception signal for camera and robotics workflows
Pairs naturally with object detection and scene captioning

Use Cases on Mixpeek

Warehouse safety agents that find near misses, blocked aisles, or unusual motion

Robotics monitoring systems that flag unexpected scene changes

Manufacturing QA review over camera feeds and process videos

Security workflows that prioritize abnormal clips before human review

Benchmarks

Dataset	Metric	Score	Source
Vad-Reasoning test set	Top-1 Hit Rate	46.44%	NVIDIA model card, 2026
Vad-Reasoning test set	Top-5 Hit Rate	83.71%	NVIDIA model card, 2026
Kinetics-400 validation	Top-1 Accuracy	85.18%	NVIDIA model card, 2026