# NVIDIA/Cosmos-Tokenizer

A suite of image and video neural tokenizers

Repository: https://github.com/NVIDIA/Cosmos-Tokenizer
Canonical: https://ross.abutalabs.com/products/cosmos-tokenizer
Homepage: https://research.nvidia.com/labs/dir/cosmos-tokenizer
Language: Jupyter Notebook
License: Apache-2.0
License Family: permissive
Topics: diffusion, tokenization, transformers
Archived: true
Last push: 2025-02-11T05:49:24+00:00

## Health v2 (maintenance only)
Score: 10/100 (v2, computed 2026-09-03T02:20:16.233290+00:00)
- activity 6, release rhythm 35, longevity 48
- inputs: {"age_days": 672, "days_push": 568, "days_rel": null, "gap_med": null, "n_releases_24m": 0}
- flags: no_releases, archived
- formula: round(0.45*activity + 0.35*rhythm + 0.20*longevity); archived -> min(score, 10)

## Adoption (not part of the score)
Stars 1731, forks 94 (observed 2026-08-28T04:05:28.682316+00:00)

## What it is
NVIDIA Cosmos Tokenizer is a suite of neural tokenizers for images and videos that convert visual data into continuous latents or discrete tokens for use in diffusion or auto-regressive transformer models. It achieves high spatial and temporal compression (up to 2048x total) and serves as a core component of the NVIDIA Cosmos video foundation model platform.

## Use cases
- tokenize videos into discrete tokens for autoregressive transformer training
- compress images into continuous latents for diffusion model training
- build physical AI or world model pipelines with efficient visual tokenization
- reduce video data storage and compute costs with high compression tokenization
- preprocess visual datasets for large multimodal generative AI models

## When to choose
- you need state-of-the-art image/video tokenization for training diffusion or autoregressive generative models
- you are building on the NVIDIA Cosmos video foundation model ecosystem
- you need extreme compression of visual data (up to 2048x) for scalable model training

## When to avoid
- you need a general-purpose image/video codec for playback or archival rather than ML tokenization
- you lack GPU resources, as the models require GPU inference
- you need active development here - the repo is read-only and moved to NVIDIA/Cosmos

## Facets
- artifact type: library
- maturity: maintenance
- function: machine-learning, deep-learning, image-processing, video-processing, serialization
- domain: deep-learning, machine-learning, computer-vision, image-processing, artificial-intelligence
- platform: python, cross-platform
- tags: tokenization, diffusion-models, transformers, video-foundation-models, latent-compression, nvidia, physical-ai, video, gpu, linux

## Member repositories
- NVIDIA/Cosmos-Tokenizer (main) score 10

## Provenance
- Observed fields: from GitHub, fetched 2026-08-28T04:05:28.682316+00:00.
- Health v2: computed from the inputs above; adoption is never an input.
- Inferred fields (summary, facets, guidance): AI-extracted, prompt v1, taxonomy v1, on 2026-08-30T03:31:41.226120+00:00, confidence not recorded.
  - readme: https://github.com/NVIDIA/Cosmos-Tokenizer (fetched 2026-08-28T04:05:28.682316+00:00, sha 3b2d7d50355e)
  - homepage: https://research.nvidia.com/labs/dir/cosmos-tokenizer (fetched 2026-08-29T11:08:25.467193+00:00, sha f700d88d9bdb)
- Data as of 2026-08-30T08:39:29.467469+00:00.