NVIDIA/cosmos-curator v2.0.0

Release v2.0.0

Repository: NVIDIA/cosmos-curator

Tag: v2.0.0

Published: 2026-05-27T20:42:29Z

Prerelease: no

Release notes:

Breaking Changes

• Rename the project from Cosmos-Curate to Cosmos Curator:
• Python imports and package paths move from cosmos_curate to cosmos_curator.
• The command-line entry point moves from cosmos-curate to cosmos-curator.
• The Helm chart path/name moves from charts/cosmos-curate to charts/cosmos-curator.
• Examples, docs, and config paths now use cosmos_curator naming, including

~/.config/cosmos_curator/config.yaml.

• Clean up split-pipeline CLI and config arguments:
• Replace --enable-sam3 with --sam3 / --no-sam3.
• Replace --enable-event-captioning with --event-captioning / --no-event-captioning.
• Replace --generate-cosmos-predict-dataset predict2 with the boolean

--generate-cosmos-predict-dataset; JSON/YAML configs now use true or false.

• Replace --artificial-text-filter enable|disable with

--artificial-text-filter / --no-artificial-text-filter.

• Replace --video-classifier enable|disable with --video-classifier /

--no-video-classifier.

• Remove the --qwen-filter-* and --qwen-video-classifier-* aliases; use

--vlm-filter-* and --video-classifier-*.

• Remove the unsupported --qwen-use-async-engine flag.
• Rename all-caps documentation filenames to lowercase kebab-case; update external links that

target old branch-relative docs paths such as docs/client/END_USER_GUIDE.md.

Added

• Qwen3.6-27B (BF16 and FP8) support for video and image captioning, registered as

qwen3_6_27b and qwen3_6_27b_fp8 variants.

• Ray Data support for Qwen captioning and TransNetV2 splitting.
• Split-output comparison tooling for summaries, captions, motion scores, and aesthetic scores.
• Run-level caption quality statistics for split-video outputs and heuristic caption quality flags.
• Video pixel budget override support for windowed vLLM captions.
• vLLM async and OpenAI backend support for SAM3 per-event captioning.
• Sensor-library overlap support, motion-vector data on CameraSensor, and a

decoder-utils-compatible sampling grid.

• Interactive Slurm launch workflow.
• Pixi development tasks for linting, CPU tests, and the cosmos-curator CLI.

Fixed

• Stabilize GPU stage autoscaling and upgrade cosmos-xenna to v0.4.2.
• Release GPU memory on stage teardown to prevent lingering CUDA contexts.
• Include the caption window end frame in CPU sampling and average caption tokens by window.
• Fix Slurm submit launches during pixi solves and guard nvidia-smi calls on CPU-only nodes.
• Repair client wheel packaging, including required storage utilities and clip-viewer assets.
• Serve marked.min.js locally in the clip viewer to avoid a CDN dependency.
• Avoid clobbering generated Dockerfiles during parallel CI builds.
• Pin PyNvVideoCodec to >=2.0.4,8.1.

Documentation

• Add split-pipeline stage reference documentation.
• Refresh captioning contracts and metadata guidance.
• Add split-output comparison and Orca agentic orchestration design documents.
• Normalize documentation filenames to lowercase kebab-case.
• Clarify host CLI versus runtime container usage.
• Update MR description guidance.