v1.5.4

What's Changed

• fix: use explicit bitrate instead of global_quality for NVENC on Blackwell by @ggorostiaga in #231

Full Changelog: 1.5.3...v1.5.4

v1.5.3

• Upgrade to Python 3.13.

Release 1.5.2

patch CT2.5 inference pipeline for single-frame JPG input

Release 1.5.1

Minor bug fixes and cleanups.

v1.5.0

1. VDR improvements
2. Multicontrol bug fix
3. Enable low latency robotics (edge deployment) inference through Transfer2.5 Distilled Edge model and inference. More distilled controlnets coming soon.

v1.4.2

• Fix bugs and clarify docs for Transfer2.5 singleview post-training
• Breakup and simplify checkpoint_db.py

v1.4.1

• Adds support for image-to-image and image-prompt inference (docs/inference_image.md)

v1.4.0

• Updated checkpoints for Cosmos-Transfer2.5-2B blur, depth, seg and edge
• Bugfix: autoregressive multiview with num_conditional_frames==0
• Optimized rendering for control video
• Updated documentation
• Added post training example for single-view edge/depth/seg/blur modalities

v1.3.3

• Fix help menu and override cli args.
• Guardrail offload fix.
• Blackwell + ARM inference support.
• Auto multiview quality fixes.
• QA fixes.
• Enable LFS for assets.

v1.3.2

• Multiview refactoring and updated checkpoints