NVIDIA/cosmos-curator v1.3.0

Release v1.3.0

Repository: NVIDIA/cosmos-curator

Tag: v1.3.0

Published: 2026-04-28T00:09:11Z

Prerelease: no

Release notes:

Added

• Image curation pipeline with semantic filtering
• Image embedding stages (Cosmos-Embed1, InternVideo2-MM, OpenAI-compatible) and image annotate pipeline
• OpenAI- and Gemini-compatible endpoints for image captioning, filtering, and classification
• Artificial-text detection stage for the video filtering pipeline (PaddleOCR-based)
• Sensor library (camera-only) with SensorGroup, mcap-based ingestion, and timestamp validation
• SeedVR-based upscaling stage
• Pipeline config files with NVCF-compatible JSON and YAML loading (--config for split/shard/dedup)
• Centralized pipeline argument validation via common_pipeline_settings and shard_pipeline_settings
• vLLM async captioning stage for higher captioning throughput (experimental — correctness

issues are still being worked through; not recommended for production use)

• OpenTelemetry instrumentation for vLLM captioning
• Token-counting instrumentation to measure captioning throughput
• Caption status fields normalized across caption backends, with status-gated metadata writing
• Stage-replay validation that compares re-run output against the original recording
• S3 support for stage-save and stage-replay
• Ray Data hello-world pipeline and splitting pipeline MVP as an alternative engine alongside Xenna
• --*-cpus-per-worker knobs documented for CPU-constrained hosts
• Run local-launched container as the host user (including AD/SSSD/NIS UIDs) to avoid root-owned outputs
• Slim Docker image built alongside the full image, with auto-warmup honoring --envs
• Local Xenna build path in CI and per-pipeline Xenna overrides
• Fixed-stride coverage in the NVCF split benchmark matrix
• Real-inference smoke test for vLLM captioning health
• Upgrade to CUDA 13.0
• Upgrade vLLM to 0.19.0
• Upgrade Ray to 2.55.0 (with the serve extra)
• Upgrade cosmos-xenna to 0.2.3
• Bump av to >=17,<18 and add the mcap dependency for the sensor library

Fixed

• SamplingGrid produced incorrect windows for irregular grids
• --execution-mode CLI flag is now honored end-to-end
• Cosmos-Embed1 writes per-variant embedding directories
• Symlink the host pixi path so shebangs resolve inside the local-launched container
• Sensor library uses read-only views to avoid accidental buffer mutation
• Add Qwen3 preprocessing logic for filtering stages
• Use pre-built images for benchmark runs to avoid redundant builds
• Remove external storage dependency from ImageSensor
• Semantic filter updates and dedup pipeline input path cleanup
• Loosen Cosmos-Reason1 caption similarity threshold to reduce flakiness

Changed

• Replace CurationPhase / PipelineBuilder with factory functions (*_builders.py); the

phase_interface module and per-pipeline phases.py files are removed

• Add config: VllmConfig parameter to VllmPlugin.make_llm_input for image vs video

modality selection; subclasses must update their signature

• Switch CI Slurm and k8s GPU jobs to the slim image with in-container pixi install and

pixi run --as-is

• Change CI NVCF backend
• Normalize the SamplingGrid API and make sampling windows explicit (no sentinel boundaries)
• Update semantic filter stages to use VllmCaptioning
• Add a CPU-only Paddle option for the unified env
• Pixi lockfile refreshed for CVE coverage
• Add notice and disclaimer to README and Docker image

Documentation

• Speed-of-light design doc for captioning throughput, with refined SOL baseline methodology

using vllm bench as the reference

• Refined Ray Data runner design with the first implementation slice
• Document --*-cpus-per-worker tuning knobs
• Add --squash-before-merge to MR guidelines