WorldFoundry
English中文

LongVie

Method overview: Framework of LongVie 2. LongVie 2 serves as a controllable video world model that integrates both dense and sparse control signals to provide world-level guidance for enhanced controllability. A degradation-aware.

LongVie from Shanghai AI Laboratory (Vchitect): image-to-video and long-video-generation; Runner parity pending; dedicated `worldfoundry-longvie-cu128`.

Edit this pageGitHubMarkdown

LongVie is a controllable ultra-long video generation framework from the Vchitect team: it drives a Wan-based diffusion backbone with dense depth and sparse point/track control videos to keep minute-scale videos consistent. This entry retains the first-generation adapter, but upstream has not published first-generation weights; the public Vchitect/LongVie2 checkpoint is not interchangeable.

Method overview: Framework of LongVie 2. LongVie 2 serves as a controllable video world model that integrates both dense and sparse control signals to provide world-level guidance for enhanced controllability. A degradation-aware.
Method overview: Framework of LongVie 2. LongVie 2 serves as a controllable video world model that integrates both dense and sparse control signals to provide world-level guidance for enhanced controllability. A degradation-aware.

The recorded paper is arXiv 2512.13604.

Project page: vchitect.github.io/LongVie2-project. Code: Vchitect/LongVie.

Build your run

Pipeline
LongVie1Pipelineworldfoundry.pipelines.longvie.pipeline_longvie
TaskInteractive Video interactive-video
Environmentworldfoundry-longvie-cu128
DeviceCUDA 12.8
Runnerworldfoundry.pipeline
Loadingfrom_pretrained
Invocationnative_or_standard_invocation
worldfoundry-eval run \
  longvie-1 \
  --pipeline.task-profile interactive-video \
  --json

What it is for

longvie-1 is the recorded depth route. Start here when you need depth maps from the documented inputs.

LongVie-1 = canonical Wan model + encoders + VAE + VACE/control roles + checkpoint loaders + the shared staged runner. The native runner preserves the official 640x352 segment size and the history/noise continuation state, with optional USP ring/ulysses distributed controls. Its synthesis package is a thin product assembly boundary, not a separate engine. It is published by Shanghai AI Laboratory (Vchitect).

Catalog tasks include image-to-video, long-video-generation, depth-controlled-video, and trajectory-controlled-video. Catalog aliases include longvie, longvie1, and longvie-v1.

Recorded evaluation entry points include ChronoMagic-Bench (Long-horizon suite for minute-scale generation).

LongVie-1 long I2V with optional depth/trajectory control. v2 is a sibling.

Runtime contract

Optional prompt, negative_prompt, input_path, interactions, and frames; artifacts generated_video and longvie-1.mp4.

CLI task profiles: interactive-video. WorldFoundry binds LongVie1Pipeline.

Family status: Runner parity pending.

Default calls fail fast unless execute=True. Execution additionally requires an explicitly staged first-generation control checkpoint because no official LongVie-1 weights are published. For the public checkpoint-backed route, use longvie-2.

WorldFoundry implements LongVie over the canonical Wan model, encoders, VAE, VACE/control roles, checkpoint loaders, and shared staged runner.

Running in WorldFoundry

Launch worldfoundry-eval run longvie-1 --pipeline.task-profile interactive-video to write depth. The command builder above switches --pipeline.task-profile among recorded tasks and still emits worldfoundry-eval run.

It runs in the dedicated environment worldfoundry-longvie-cu128 (Python 3.10, CUDA 12.8).

Recorded weight repositories include Wan-AI/Wan2.1-I2V-14B-480P and Wan-AI/Wan2.1-T2V-1.3B. Prepare the environment and check local assets with bash scripts/inference/prepare_model_infer.sh longvie-1. Add --download to fetch missing public checkpoints.

Status: Runner parity pending.