WorldFoundry
English中文

SkyReels-V2

Method overview: Overview of the proposed method.

SkyReels-V2 from Skywork AI (Kunlun): text-to-video and image-to-video; Runner parity pending.

Edit this pageGitHubMarkdown

SkyReels-V2 is Skywork AI's open film-generation model family built on the Wan architecture, recorded here for text-to-video and image-to-video. WorldFoundry serves the 1.3B route assembled entirely from the shared native Wan graph, conditioner, VAE, scheduler, initializer, loader, and runner — the vendored CLI/pipeline is deliberately not an inference backend.

Official paper teaser: SkyReels-V2 produces stunningly realistic and cinematic high-resolution videos of virtually unlimited length. The model excels at maintaining visual consistency of the main subject across all frames, ensuring no.
Official paper teaser: SkyReels-V2 produces stunningly realistic and cinematic high-resolution videos of virtually unlimited length. The model excels at maintaining visual consistency of the main subject across all frames, ensuring no.
Method overview: Overview of the proposed method.
Method overview: Overview of the proposed method.

The recorded paper is arXiv 2504.13074.

Code: SkyworkAI/SkyReels-V2.

Build your run

Pipeline
SkyReelsV2Pipelineworldfoundry.pipelines.skyreels_v2.pipeline_skyreels_v2
TaskInteractive Video interactive-video
Environmentworldfoundry-unified-cu128
DeviceCUDA 12.8
Runnerworldfoundry.pipeline
Loadingfrom_pretrained
Invocationnative_or_standard_invocation
worldfoundry-eval run \
  skyreels-v2 \
  --pipeline.task-profile interactive-video \
  --json

What it is for

Pick skyreels-v2 when one recipe covers text-to-video and image-to-video: a prompt in, an optional still, a clip out. Start here instead of wiring two upstream demos.

The manifest records the upstream graph as an FPS-conditioned Wan capability: SkyReels-V2's diffusion-forcing design conditions on frame rate to enable long-form generation. In WorldFoundry it is a native Wan-graph assembly with prompt/image inputs and a generated_video artifact. It is published by Skywork AI (Kunlun).

Catalog tasks include text-to-video and image-to-video. Documented modalities: text and image in, video out.

Recorded evaluation entry points include VBench (Standard T2V quality suite).

Not SkyReels V3. Score on VBench.

Runtime contract

Optional prompt, negative_prompt, frames, fps, and height; artifacts skyreels-v2.mp4.

CLI task profiles: interactive-video. WorldFoundry binds SkyReelsV2Pipeline.

Family status: Runner parity pending.

Integration is complete for the 1.3B route; no runner-parity record exists yet. Larger upstream checkpoints (5B/14B) are not the recorded route — stick to 1.3B for the supported path.

The upstream graph is represented as an FPS-conditioned Wan capability; the vendored CLI/pipeline is not an inference backend.

Running in WorldFoundry

The catalog launch is worldfoundry-eval run skyreels-v2 --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 Skywork/SkyReels-V2-DF-1.3B-540P, Skywork/SkyReels-V2-T2V-14B-720P, and Skywork/SkyReels-V2-I2V-1.3B-540P among others. Prepare the environment and check local assets with bash scripts/inference/prepare_model_infer.sh skyreels-v2. Add --download to fetch missing public checkpoints.

Recorded licenses include other; status: Runner parity pending.