WorldFoundry
English中文

Kairos Sensenova

Method overview: Framework of Kairos. World Understanding extracts a control-sufficient state Z t Z_{t} ; World Generation regularizes physical consistency of Z t Z_{t} through future imagination; World Prediction uses Z t Z_{t} for.

Kairos Sensenova from SenseTime (SenseNova): text-to-video and image-to-video; Verified; dedicated `kairos-sensenova`.

Edit this pageGitHubMarkdown

Kairos Sensenova is SenseTime's embodied world-model video generator: a 4B model covering text-, image-, and text+image-to-video with a robot-oriented distilled recipe. In this catalog both integration and runner are verified — the default robot 4B 480P DMD text-to-video recipe runs through the in-tree Apache-2.0 torchrun runtime with 8 ranks.

Official paper teaser: Motivation of Kairos. Existing world models have advanced along representational, generative, interactive, and unified world-action directions, providing useful capabilities such as abstract reasoning, high-fidelity.
Official paper teaser: Motivation of Kairos. Existing world models have advanced along representational, generative, interactive, and unified world-action directions, providing useful capabilities such as abstract reasoning, high-fidelity.
Method overview: Framework of Kairos. World Understanding extracts a control-sufficient state Z t Z_{t} ; World Generation regularizes physical consistency of Z t Z_{t} through future imagination; World Prediction uses Z t Z_{t} for.
Method overview: Framework of Kairos. World Understanding extracts a control-sufficient state Z t Z_{t} ; World Generation regularizes physical consistency of Z t Z_{t} through future imagination; World Prediction uses Z t Z_{t} for.

The recorded paper is arXiv 2606.16533.

Code: kairos-agi/kairos-sensenova.

Build your run

Pipeline
KairosPipelineworldfoundry.pipelines.kairos.pipeline_kairos
TaskDefault default
Environmentkairos-sensenova
DeviceCUDA 12.8
Runnerworldfoundry.pipeline
Loadingfrom_pretrained
Invocationnative_or_standard_invocation
worldfoundry-eval run \
  kairos-sensenova \
  --pipeline.task-profile default \
  --json

What it is for

kairos-sensenova is the recorded interactive world line. Start here when the job is stepping a world with actions.

The default runnable variant follows the official robot 4B 480P distilled recipe (kairos_4b_config_DMD.py): 81 frames, 480x640, cfg_scale 5, tiled VAE decode, no CFG-parallel. Recorded checkpoints cover the robot 480P distilled, 480P pretrained, and 720P variants; the config layout expects Kairos, Qwen, and Wan2.1-T2V-14B subdirectories. It is published by SenseTime (SenseNova).

Catalog tasks include text-to-video, image-to-video, text-image-to-video, and embodied-world-model. Catalog aliases include Kairos, Kairos Sensenova, kairos-4b, and kairos-sensenova-4b. Documented modalities: text and image in, video out.

Recorded evaluation entry points include VBench (Video quality for text/image-to-video output), EWMBench (Embodied world-model prediction quality).

Runtime contract

Optional prompt, negative_prompt, frames, height, and width; artifacts kairos-sensenova.mp4.

CLI task profiles: default. WorldFoundry binds KairosPipeline.

Family status: Verified.

Recorded default defaults: frames=81, height=480, width=640, seed=0.

Set WORLDFOUNDRY_KAIROS_MODELS_ROOT (or models_root) to a directory matching the official config layout. 720P and non-distilled variants need the corresponding checkpoint paths and config override.

Running in WorldFoundry

Launch worldfoundry-eval run kairos-sensenova --pipeline.task-profile default to step the world. The command builder above switches --pipeline.task-profile among recorded tasks and still emits worldfoundry-eval run.

Recorded weight repositories include kairos-agi/kairos-sensenova-robot-4B-480P-distilled, kairos-agi/kairos-sensenova-4B-480P-pretrained, and kairos-agi/kairos-sensenova-4B-720P. Prepare the environment and check local assets with bash scripts/inference/prepare_model_infer.sh kairos-sensenova. Add --download to fetch missing public checkpoints.

Status: Verified.