WorldFoundry
English中文

SkyReels V3

SkyReels V3

SkyReels V3 from Skywork AI (Kunlun): reference-to-video and audio-guided-video; Runner parity pending.

Edit this pageGitHubMarkdown

SkyReels V3 is Skywork's newer generation, recorded for reference-to-video, audio-guided video, and video-to-video. The catalog route loads the SkyReels-V3-R2V-14B checkpoint into the shared native Wan graph with reference-image latent concatenation and dual text/image guidance — implemented as reusable initializer and execution-strategy capabilities rather than a bespoke runtime.

Official paper teaser: Reference Images to Video Results. SkyReels-V3 can facilitate dynamic interplay between different subjects within specified contexts.
Official paper teaser: Reference Images to Video Results. SkyReels-V3 can facilitate dynamic interplay between different subjects within specified contexts.

The recorded paper is arXiv 2601.17323.

Code: SkyworkAI/SkyReels-V3.

Build your run

Pipeline
SkyReelsV3Pipelineworldfoundry.pipelines.skyreels_v3.pipeline_skyreels_v3
TaskInteractive Video interactive-video
Environmentworldfoundry-unified-cu128
DeviceCUDA 12.8
Runnerworldfoundry.pipeline
Loadingfrom_pretrained
Invocationnative_or_standard_invocation
worldfoundry-eval run \
  skyreels-v3 \
  --pipeline.task-profile interactive-video \
  --json

What it is for

skyreels-v3 is the recorded prompt-to-video route — not image editing and not a policy. Start here when the job is a clip from text.

Transformer, UMT5 text encoder, and VAE checkpoint schemas map exactly onto the native Wan graphs (recorded as verified). Reference conditioning concatenates reference-image latents; guidance runs dual text/image channels through the UniPC scheduler and dual-condition runner. It is published by Skywork AI (Kunlun).

Catalog tasks include reference-to-video, audio-guided-video, video-to-video, and video-generation. Documented modalities: image and video in, video out.

Recorded evaluation entry points include VBench (Baseline quality suite).

Not SkyReels V2. Score on VBench.

Runtime contract

Optional prompt, negative_prompt, input_path, frames, and fps; artifacts skyreels-v3.mp4.

CLI task profiles: interactive-video. WorldFoundry binds SkyReelsV3Pipeline.

Family status: Runner parity pending.

Runner status is partial: checkpoint schema mapping is exact, and a prior official-runtime visual artifact is documented, but native visual parity still needs a GPU inference run. No Diffusers or model-owned subprocess runtime is involved.

WorldFoundry loads the SkyReels V3 checkpoint into the shared native Wan graph and framework-owned diffusion runner.

Running in WorldFoundry

Launch worldfoundry-eval run skyreels-v3 --pipeline.task-profile interactive-video to sample the clip. 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-V3-R2V-14B. Prepare the environment and check local assets with bash scripts/inference/prepare_model_infer.sh skyreels-v3. Add --download to fetch missing public checkpoints.

Status: Runner parity pending.