---
title: "DualCamCtrl"
description: "HKUST-GZ EnVision: RGB + depth + camera .torch. Official demo verified: 320×480, 61 frames, 50 steps. License unknown. Expected seaside.mp4."
modelId: "dualcamctrl"
pageSource: authored
---

DualCamCtrl steers video with a **camera trajectory plus a depth image**: RGB and depth together pin geometry so the camera move stays grounded. WorldFoundry vendors infer under `worldfoundry/synthesis/visual_generation/dualcamctrl` as `DualCamCtrlPipeline`. Runner **and** demo are **verified** against the official RGB / depth / camera contract.

<ModelPaperFigures modelId="dualcamctrl" locale="en" />

Paper: [arXiv 2511.23127](https://arxiv.org/abs/2511.23127). Project: [soyouthinkyoucantell.github.io/dualcamctrl-page](https://soyouthinkyoucantell.github.io/dualcamctrl-page/). Code: [EnVision-Research/DualCamCtrl](https://github.com/EnVision-Research/DualCamCtrl). Publisher: HKUST (Guangzhou) EnVision Lab. Catalog license: **unknown**.

<ModelCommandBuilder modelId="dualcamctrl" locale="en" />

## What it is for

Dual-camera control for a stereo-style pair.

Launch `worldfoundry-eval run dualcamctrl`.

Score on [CameraBench](/docs/evaluation/benchmark-hub/camerabench) — camera-trajectory adherence of controlled generation.

## Runtime contract

Required: `prompt` and `image`; optional: `video`; `actions`: camera_trajectory, depth_image; artifacts: `dualcamctrl.mp4` and `seaside.mp4`.

| Knob | Official demo |
| --- | --- |
| Size | **320×480** |
| Frames | **61** |
| Steps | **50** |
| Trajectory | `.torch` camera file |

## Run in WorldFoundry

Status: **Verified**. Unified environment `worldfoundry-unified-cu128` (Python 3.11, CUDA 12.8). Stage `FayeHongfeiZhang/DualCamCtrl`, `alibaba-pai/Wan2.1-Fun-V1.1-1.3B-Control-Camera`, and `Wan-AI/Wan2.1-T2V-1.3B` among others locally — missing files fail at load, nothing is downloaded. Run `worldfoundry-eval run dualcamctrl`. Catalog VRAM is unset.

<ModelRelatedRecipes modelId="dualcamctrl" locale="en" />