Skip to content

Changelog

All notable changes to strands-cosmos are documented here.

The format is based on Keep a Changelog, and this project adheres to Semantic Versioning.

Unreleased

[0.4.4] - 2026-06-04

Changed

  • cosmos3_video2video now exposes the conditioning controls that actually drive how much the prompt transforms the video: condition_frames (latent frame indexes kept clean) and condition_keep ("first"/"last"). New restyle-friendly defaults (condition_frames="0", condition_keep="last", guidance=8, steps=35) so a real transform (day→night, recolor, change the scene) is the easy path. With the old defaults the prompt barely changed the video (it reconstructed the input).
  • c3-v2v justfile recipe gains matching cond_frames / cond_keep params.
  • Added demo/cosmos3_v2v/ showcase (recolor / rain / crowd / night) + tuning guide.

[0.4.3] - 2026-06-04

Added — native video-to-video transfer

  • cosmos3_video2video tool + c3-omni-docker / c3-v2v justfile recipes: structure-preserving video transfer (e.g. day→night, recolor buildings, restyle) by re-rendering an input video with a new prompt via the Cosmos 3 vLLM-Omni server (/v1/videos/sync, input_reference).
  • vLLM-Omni runs from the official vllm/vllm-omni:cosmos3 Docker image (the only build with all modalities incl. video2video). c3-omni-docker launches it with --gpus all and the shared HF cache.
  • Validated live: input construction-site video transformed to night / recolor / rain (832×480, structure preserved, ~12s/clip).

[0.4.2] - 2026-06-04

Fixed

  • All justfile-backed tools failed on pip install: the justfile (which every cosmos_* / cosmos3_* tool shells out to) was never packaged, so _find_justfile() returned None in a pip install and tools errored with exit 127. The justfile is now bundled into the wheel/sdist at strands_cosmos/justfile (copied from the repo root at build time) and the lookup resolves it. Verified from a clean wheel install. Added a regression test.

[0.4.1] - 2026-06-04

Fixed

  • PyPI publish failure (0.3.2 / 0.4.0 never reached PyPI): the cosmos3-gen and all extras declared a direct-URL dependency (diffusers @ git+https://...), which PyPI rejects on upload. The extras now require plain diffusers>=0.36; the Cosmos3OmniPipeline dev build is installed separately via just c3-setup-gen or pip install -U "git+https://github.com/huggingface/diffusers.git".
  • Release workflow now fails fast if built metadata contains any direct-URL dependency (guard step before PyPI publish).

[0.4.0] - 2026-06-04

Added — Cosmos 3 post-training (SFT)

  • 7 cosmos3_train_* tools + c3-train-* justfile recipes wrapping the NVIDIA Cosmos Framework supervised fine-tuning stack: cosmos3_train_recipes, cosmos3_train_show, cosmos3_train_convert, cosmos3_train_convert_vlm, cosmos3_train_prep_dataset, cosmos3_train, cosmos3_train_export.
  • Full 4-step flow: checkpoint → DCP convert, dataset prep, SFT run (paired launch shell / torchrun -m cosmos_framework.scripts.train), and DCP → HF safetensors export. Recipes: vision_sft_nano/super, llava_ov, videophy2_nano.
  • New Cosmos 3 Training guide + examples/10_cosmos3_finetune.py; c3-doctor reports the training env.
  • Validated locally (no 8× H100 needed): c3-train-recipes lists recipes and c3-train-show resolves/validates the full SFT config via train.py --dryrun.

[0.3.2] - 2026-06-04

Added

  • cosmos3-gen optional extra — pip install "strands-cosmos[cosmos3-gen]" installs the Cosmos 3 generator backend in one step (Diffusers dev build, cosmos_guardrail, soundfile, imageio), making text/image -> image/video/sound generation frictionless without the justfile.

Fixed

  • just c3-setup-gen now also installs soundfile (required for the video-with-sound audio mux path).

[0.3.1] - 2026-06-04

Fixed

  • Video decoding on fresh installs: transformers 5.x decodes video via torchcodec and silently falls back to torchvision, whose io.read_video was removed in torchvision>=0.27 (AttributeError when captioning a video). torchcodec is now a core dependency so video works out of the box.

0.3.0 - 2026-06-04

Added — Cosmos 3 omnimodal world models 🌌

First-class support for NVIDIA's newest Cosmos 3 omnimodal model family — reasoning and generation across text, image, video, audio, and action. Runs on local compute (vLLM / Diffusers / Cosmos Framework).

  • Model providers (implement strands.models.Model):
  • Cosmos3ReasonerModel — omnimodal reasoning (text + vision → text) via a local vLLM server (Cosmos3ReasonerForConditionalGeneration). Captioning, temporal localization, embodied next-action, 2D grounding, physical plausibility, situation understanding, action chain-of-thought. Supports explicit <think> reasoning.
  • Cosmos3GeneratorModel — generation (text/image → image/video/sound) in-process via HuggingFace Diffusers Cosmos3OmniPipeline. Generates and muxes stereo AAC audio @ 48kHz for video-with-sound.
  • 16 cosmos3_* tools (thin justfile wrappers):
  • Reasoner: cosmos3_reason, cosmos3_caption, cosmos3_temporal, cosmos3_embodied, cosmos3_ground, cosmos3_plausibility, cosmos3_situation, cosmos3_action_cot
  • Generator: cosmos3_text2image, cosmos3_text2video, cosmos3_image2video, cosmos3_text2video_sound
  • Action / world-model: cosmos3_forward_dynamics, cosmos3_inverse_dynamics, cosmos3_policy
  • Servers: cosmos3_serve
  • 13 c3-* justfile recipes: c3-doctor, c3-setup-{reason,gen,omni,framework}, c3-serve-{reason,omni,status,stop-reason,stop-omni}, c3-reason, c3-gen, c3-action.
  • cosmos3 optional extra (pip install "strands-cosmos[cosmos3]") for the reasoner OpenAI client.
  • Examples: 06_cosmos3_reason.py, 07_cosmos3_generate.py, 08_cosmos3_action.py, 09_cosmos3_showcase.py (reason → generate showcase).
  • Showcase: demo/cosmos3_showcase/ — Cosmos 3 reasons about a real video, then generates similar videos (incl. one with synchronized audio) from its own description.
  • Docs: new Cosmos 3 Guide; README, index, quickstart, installation, architecture, API reference, and capability guides all lead with Cosmos 3.

Changed

  • README and docs now present 4 model providers + 37 tools and lead with Cosmos 3 as the flagship; Cosmos-Reason2 repositioned as the lightweight edge/Jetson VLM.
  • c3-doctor recognizes cached Hugging Face tokens (not just HF_TOKEN/env).
  • c3-serve-status (and cosmos3_serve) detect directly-launched servers via an HTTP /health probe, not only the PID file.

Verified (single NVIDIA L40S, 46 GB, no NIM)

  • Reasoner: caption 6.6s, plus temporal / embodied / plausibility / situation.
  • Generator: text→image, text→video, image→video, and text→video + sound (H264 + AAC stereo 48kHz).
  • Action: forward-dynamics rollout (832×480, 61 frames) via the Cosmos Framework.

Notes

  • CUDA pairing: match the torch backend to your driver — CUDA 13 → cu130 + vllm==0.21.0; CUDA 12.8 → cu128 + vllm==0.19.1. just c3-doctor reports it.
  • Single-GPU: the reasoner (vLLM) and generator (Diffusers) each load a 16B model and will not fit on one ~46 GB GPU simultaneously — stop one before running the other.
  • The Cosmos 3 reasoner caps --max-model-len 32768 to avoid KV-cache OOM (the model's 262K default exceeds a single 46 GB GPU).
  • Removed the internal COSMOS3_INTEGRATION.md planning doc (superseded by shipped code + docs).

0.2.0 - 2026-05-08

Added

  • 21 tools covering the full Cosmos pipeline (Reason2 inference, Predict2.5, Transfer2.5, model lifecycle, training, distillation, Xenna curation, evaluation, I/O, system diagnostics) — all thin wrappers over a 50+ recipe justfile.
  • MkDocs Material documentation site with examples, guides, and API reference.
  • Notebooks and runnable examples (01–05).
  • AGENTS.md development contract.

0.1.2 - 2026-03-xx

Fixed

  • asyncio event-loop thread-safety in cosmos_vision_invoke.

0.1.1 - 2026-03-09

Added

  • Initial public release: CosmosVisionModel (video + image + text) and CosmosModel (text-only) Strands model providers for Cosmos-Reason2.
  • Jetson CUBLAS compatibility fix (strands-cosmos-fix-cublas).
  • Dashcam safety-analysis demo.