v0.2.0

Changes since v0.1.0.

Added

• Cosmos3 Nano Sim Auto:
  interactive and batch driving with selectable RGB cameras, optional joint
  LiDAR generation, and automatically cached model weights and prepared
  residential, rain, and snow scenes.
• Cosmos3 Nano Sim Bimanual
  and HandDreams: action-conditioned robotics video using the public GA checkpoint,
  Quest/Isaac Teleop or relayed controller input, and controller-free recorded-action replay.
• A v2 Demo API with application/session
  lifecycle management, separate model and UI loops, and WebRTC, native-window,
  and MP4 output. Custom presentation loops, multiview conditioning, and sensor
  displays are reusable; driving adapters can supply cameras, rendering backends,
  and scene loaders.
• Accelerated inference components
  for quantization, streaming attention, and bounded KV caches. Tensor/context
  parallel runtime support is available to compatible model adapters.
• Waypoint 1.5 1B, SANA-WM, and SwiftVR integrations; LingBot-World v2 support.
• Profiling with IProfiler,
  NVTX ranges, stage timings, frame rates, and input latency; headless model benchmarks.
• Offline executable packaging for installed
  v2 applications on Windows and Linux, with asset preloading and preparation validation.
• Crazy Robotaxi, a standalone application with configurable controls, persistent
  settings, live prompt editing, free-roam mode, and a 32 GB RTX 5090 preset.

Changed

• Migrated demos and model adapters use apps/ and integrations_v2/.
  Launch them with flashdreams-run-v2; put runtime options before -- and
  application options after it. Custom model/UI loop steps return list[StepResult].
  See the application guide.
• V2 MP4 runs require an explicit --timeout and/or --total-model-steps.
  Use --timeout unbound when the application supplies its own stopping condition.
• Added CUDA 12/13 installation groups (cuda13 by default; cuda12 on Linux)
  and deferred optional encoder imports until the encoder is used.

Removed

• HY-WorldPlay integration and legacy OmniDreams evaluator entry points.
  Use the v2 OmniDreams applications for driving and generation.

Fixed

• WebRTC latest-frame delivery and control-input latency; device-aware
  postprocessing, including multi-device Interactive Drive pipelines.
• Cosmos3 driving prompt conditioning, short control inputs during inference,
  and braking after focus loss. The demo now warns when driving leaves prepared map coverage.
• Wan 2.2 VAE channel ordering, Wan streaming decoder memory layouts, animated
  AVIF video-input selection, and decoding beyond the requested frame window.

Security

• PyTorch checkpoint loading now uses weights_only=True for local files,
  S3 downloads, and merged distributed-checkpoint caches.

Full changelog: v0.1.0...v0.2.0