WorldFoundry
English中文

WorldCam

WorldCam

WorldCam from KAIST / Adobe Research / MAUM AI: world and camera control; Verified.

Edit this pageGitHubMarkdown

WorldCam, from KAIST's CV Lab with Adobe Research and MAUM AI, is an interactive world model with camera-pose control for long-horizon game video. In this catalog the integration is verified: the runtime code is in-tree, the wrapper supports load_model=False planning without external checkout dependency, and the runner supports the official-style conditioning-video demo in the unified environment profile.

Official paper teaser: Teaser (Best viewed in color and zoomed in): WorldCam is an interactive 3D gaming model that enables precise action control under challenging keyboard and mouse inputs (top), supports long-horizon interactions.
Official paper teaser: Teaser (Best viewed in color and zoomed in): WorldCam is an interactive 3D gaming model that enables precise action control under challenging keyboard and mouse inputs (top), supports long-horizon interactions.

The recorded paper is arXiv 2603.16871.

Project page: cvlab-kaist.github.io/WorldCam. Code: cvlab-kaist/WorldCam.

Build your run

Pipeline
WorldCamPipelineworldfoundry.pipelines.worldcam.pipeline_worldcam
TaskInteractive Video interactive-video
Environmentworldfoundry-unified-cu128
DeviceCUDA 12.8
Runnerworldfoundry.pipeline
Loadingfrom_pretrained
Invocationnative_or_standard_invocation
worldfoundry-eval run \
  worldcam \
  --pipeline.task-profile interactive-video \
  --pipeline.input-path VALUE \
  --json

What it is for

Use worldcam for camera-conditioned video inference. The recipe binds the recorded camera-control profile.

The in-tree runner follows the official conditioning-video path and consumes local WorldCam/Wan assets, indicating a Wan-based diffusion backbone with camera-pose conditioning. Architecture internals beyond this are not recorded in this repository. It is published by KAIST / Adobe Research / MAUM AI.

Catalog tasks include world, camera control, and world model.

Recorded evaluation entry points include CameraBench (Camera-pose-controlled rollout adherence), WorldModelBench (Long-horizon interactive world quality).

Runtime contract

Required input_path; optional prompt, negative_prompt, fps, height, and width; artifacts worldcam.mp4.

CLI task profiles: interactive-video. WorldFoundry binds WorldCamPipeline.

Family status: Verified.

Recorded interactive-video defaults: fps=30, height=480, width=832, seed=0.

The release is code-only (no confirmed official checkpoints), so stage local WorldCam/Wan assets yourself; the wrapper's load_model=False mode lets you validate planning without weights.

Running in WorldFoundry

The catalog launch is worldfoundry-eval run worldcam --pipeline.task-profile interactive-video. The command builder above switches --pipeline.task-profile among recorded tasks and still emits worldfoundry-eval run.

It runs in the unified environment worldfoundry-unified-cu128 (Python 3.11, CUDA 12.8).

Recorded weight repositories include worldcam/worldcam and Wan-AI/Wan2.1-T2V-1.3B. Prepare the environment and check local assets with bash scripts/inference/prepare_model_infer.sh worldcam. Add --download to fetch missing public checkpoints.

Status: Verified.