Cosmos-Embed1 is a joint video-text embedder tailored for physical AI. It can be used for text-to-video retrieval, inverse video search, semantic deduplication, zero-shot and k-nearest-neighbors (kNN) classification, and as a base model for video curation tasks. It has state-of-the-art (SOTA) performance on autonomous vehicle (AV) and robotics datasets, while maintaining competitive performance in general domains. This model is ready for commercial use.
Model Developer: NVIDIA
Model Versions
The Cosmos-Embed1 release includes the following embedders:
Cosmos-Embed1
Cosmos-Embed1-224p (optimized with 8 frames and 224x224 input resolution, 256-dim output text and video embeddings)
Cosmos-Embed1-336p (optimized with 8 frames and 336x336 input resolution, 768-dim output text and video embeddings)
Cosmos-Embed1-448p (optimized with 8 frames and 448x448 input resolution, 768-dim output text and video embeddings)
Note, while each checkpoint was optimized at a specific fixed resolution (and default to these), they all support arbitrary non-square resolutions.
Under the NVIDIA Open Model License, NVIDIA confirms:
Models are commercially usable.
You are free to create and distribute Derivative Models.
NVIDIA does not claim ownership to any outputs generated using the Models or Derivative Models.
Important Note: If you bypass, disable, reduce the efficacy of, or circumvent any technical limitation, safety guardrail or
associated safety guardrail hyperparameter, encryption, security, digital rights management, or authentication mechanism contained
in the Model, your rights under NVIDIA Open Model License Agreement will automatically terminate.
Deployment Geography
Global
Use Case
Physical AI: encompassing robotics, autonomous vehicles (AV) etc.
The architecture is based on QFormer, with modifications for processing video inputs.
The video embedder processes frames individually with a ViT backbone. The per-frame ViT features are concatenated in the temporal dimension and augmented with temporal embeddings. These are then passed into the QFormer which summarizes via cross-attention a compact set of visual query tokens from the provided frames. The visual query tokens are then pooled into a single video embedding. The text embedder processes tokenized text via the self-attention branch of the Qformer to produce a text embedding.
The normalized text and video embeddings are aligned via a contrastive video-text loss, as well as auxiliary losses such as video-text matching and video captioning. For the 336p and 448p variants, we additionally use summary and dense distillation losses.
image/jpeg
Input/Output Specifications
Input
Input Type(s): Text+Video
Input Format(s):
Text: UTF-8 string
Video: tensor scaled from 0 to 1 of RGB frame sequences.
Input Parameters:
Text: One-dimensional (1D)
Video: Three-dimensional (3D)
Other Properties Related to Input:
The input string will be truncated or padded to 128 text tokens. When used for text-to-video (T2V) retrieval, it should contain a short description of the object, scene or action of interest.
Arbitrary, non-square resolutions are supported in the inference code. This can be configured at model loading time.
The model architecture supports input videos of varying lengths, but it has been optimized for 8 frames, sampled at 1-2 frames per second (FPS).
Output
Output Type(s): Text+Video
Output Format(s):
Text: floating-point normalized vector of size 256 (for 224p variant) or 768 (for 336p and 448p variants).
Video: floating-point normalized vector of size 256 (for 224p variant) or 768 (for 336p and 448p variants).
Output Parameters:
Text: One-dimensional (1D)
Video: One-dimensional (1D)
Other Properties Related to Output: Continuous-valued L2-normalized feature vectors with a dimensionality of 256 or 768. A distance can be calculated between embeddings using cosine distance. Dense intermediate feature maps are also provided for convenience.
Our AI models are designed and/or optimized to run on NVIDIA GPU-accelerated systems. By leveraging NVIDIA’s hardware (e.g. GPU cores) and software frameworks (e.g., CUDA libraries), the model achieves faster training and inference times compared to CPU-only solutions.
Note: We have only tested Cosmos-Embed1 with BF16 precision on Ampere and Hopper GPUs. If you are using older versions of NVIDIA GPUs (e.g., NVIDIA Volta GPUs), you may need to switch to FP32 precision.
Operating System(s)
Linux (We have not tested on other operating systems.)
Usage
Installation
The main model and processor dependencies can be fetched with:
pip install transformers einops torch torchvision
One can optionally install Transformer Engine for faster inference:
A code snippet for video and text inference is shown below.
For a step-by-step guide, please refer to the Juypter notebook here.
import decord
import numpy as np
import torch
from transformers import AutoProcessor, AutoModel
import subprocess
import io
# load model and pre-processor
model = AutoModel.from_pretrained("nvidia/Cosmos-Embed1-448p", trust_remote_code=True).to("cuda", dtype=torch.bfloat16)
preprocess = AutoProcessor.from_pretrained("nvidia/Cosmos-Embed1-448p", trust_remote_code=True)
# load mock data
video_url = "https://upload.wikimedia.org/wikipedia/commons/3/3d/Branko_Paukovic%2C_javelin_throw.webm"
subprocess.check_call(["wget", "-O", "/tmp/javelin_throw.mp4", video_url])
reader = decord.VideoReader("/tmp/javelin_throw.mp4")
frame_ids = np.linspace(0, len(reader)-1, 8, dtype=int).tolist()
frames = reader.get_batch(frame_ids).asnumpy()
batch = np.transpose(np.expand_dims(frames, 0), (0, 1, 4, 2, 3)) # BTCHW
captions = [
"a person riding a motorcycle in the night",
"a car overtaking a white truck",
"a video of a knight fighting with a sword",
"a man wearing red spandex throwing a javelin",
"a young man javelin throwing during the evening", # distractor
"a man throwing a javelin with both hands", # distractor
]
# video and text processing
video_inputs = preprocess(videos=batch).to("cuda", dtype=torch.bfloat16)
video_out = model.get_video_embeddings(**video_inputs)
text_inputs = preprocess(text=captions).to("cuda", dtype=torch.bfloat16)
text_out = model.get_text_embeddings(**text_inputs)
# ranking and argmax
probs = (torch.softmax(model.logit_scale.exp() * video_out.visual_proj @ text_out.text_proj.T, dim=-1))[0]
print(captions[probs.argmax()])
Training and Evaluation
We train and evaluate the Cosmos-Embed1 models on a variety of video datasets covering zero-shot classification and retrieval.
We use public reference training/test splits when available (e.g. Kinetics), otherwise we take a 90/10% split. The training pool is approximately 8m unique videos with multiple curated captions, sampled from robotics, autonomous vehicle, activity recognition and general domains.
AV action recognition (internal): Hybrid: Human, Automated
Curated captioned video dataset (internal): Automated
Metrics
We compare with the state-of-the-art text/video embedders: InternVideo2-1B and Perception Encoder. As input, we use 8 linearly spaced frames per clip and resize to square inputs.
Evaluation metrics:
(Class-Weighted) F1-score for text-video zero-shot classification
Text-to-video (T2V) and video-to-text (V2T) recall at k=1 for multi-modal retrieval, including DSL rerank.
NVIDIA believes Trustworthy AI is a shared responsibility and we have established policies and practices to enable development for a wide array of AI applications. When downloaded or used in accordance with our terms of service, developers should work with their internal model team to ensure this model meets requirements for the relevant industry and use case and addresses unforeseen product misuse.
Users are responsible for model inputs and outputs. Users are responsible for ensuring safe integration of this model, including implementing guardrails as well as other safety mechanisms, prior to deployment.
For more detailed information on ethical considerations for this model, please see the subcards of Explainability, Bias, Safety & Security, and Privacy below.
Please report security vulnerabilities or NVIDIA AI Concerns here.
Plus Plus (++) Promise
We value you, the datasets, the diversity they represent, and what we have been entrusted with. This model and its associated data have been:
Verified to comply with current applicable disclosure laws, regulations, and industry standards.
Verified to comply with applicable privacy labeling requirements.
Annotated to describe the collector/source (NVIDIA or a third-party).
Characterized for technical limitations.
Reviewed to ensure proper disclosure is accessible to, maintained for, and in compliance with NVIDIA data subjects and their requests.
Reviewed before release.
Tagged for known restrictions and potential safety implications.
Bias
Field
Response
Participation considerations from adversely impacted groups protected classes in model design and testing:
None
Measures taken to mitigate against unwanted bias:
None
Explainability
Field
Response
Intended Application & Domain:
Embedding of text and videos for physical AI
Model Type:
ViT, QFormer
Intended Users:
Physical AI developers
Output:
Text/Video embedding vectors
Describe how the model works:
Projects inputs into aligned embedding space (text/video).
Technical Limitations:
Due to the training datasets being predominantly composed of short action-focused English captions, different text prompts may not be properly aligned to video data, producing suboptimal matching.
Verified to have met prescribed NVIDIA quality standards:
The Principle of least privilege (PoLP) is applied limiting access for dataset generation and model development. Restrictions enforce dataset access during training, and dataset license constraints adhered to. Model checkpoints are made available on Hugging Face, and may become available on cloud providers' model catalog.
