WorldFoundry
English中文

Wan2.1

Method overview: Wan-Encoder and Wan-Decoder around stacked DiT blocks, with umT5 text entering by cross-attention.

Alibaba Tongyi Lab's open video family — T2V 1.3B/14B and I2V 14B at 480p/720p. The 1.3B 832×480, 81-frame route is the verified catalog backbone.

Edit this pageGitHubMarkdown

Wan2.1 is Alibaba Tongyi Lab's open video foundation family, and the backbone a large part of this catalog builds on — camera-control world models, action-conditioned rollouts, and later Helios-style long-video work all start from these weights. The official suite covers text-to-video at 1.3B and 14B, image-to-video 14B at 480p and 720p, plus upstream video editing and bilingual visual text (Chinese and English). This page records the four Apache-2.0 Hugging Face lines WorldFoundry actually binds: wan2.1-t2v-1.3b, wan2.1-t2v-14b, wan2.1-i2v-14b-480p, and wan2.1-i2v-14b-720p.

Official paper teaser: Wan2.1 text-to-video samples from the paper.
Official paper teaser: Wan2.1 text-to-video samples from the paper.
Method overview: Wan-Encoder and Wan-Decoder around stacked DiT blocks, with umT5 text entering by cross-attention.
Method overview: Wan-Encoder and Wan-Decoder around stacked DiT blocks, with umT5 text entering by cross-attention.

The technical report is Wan: Open and Advanced Large-Scale Video Generative Models. Code and weights live at Wan-Video/Wan2.1 and Wan-AI.

Build your run

Pipeline
Wan2p1T2VPipelineworldfoundry.pipelines.wan.pipeline_wan_2p1_t2v
TaskInteractive Video interactive-video
Environmentworldfoundry-unified-cu128
DeviceCUDA 12.8
Runnerworldfoundry.pipeline
Loadingfrom_pretrained
Invocationnative_diffusion_invocation
worldfoundry-eval run \
  wan2.1-t2v-1.3b \
  --pipeline.task-profile interactive-video \
  --json

Variants

  • wan2.1-t2v-1.3b
    StatusRunner parity pending

    text-to-video

    Runtime profilewan2.1-t2v-1.3b-832x480-81f

    Pipeline bindingwan2.1-t2v-1.3b

  • wan2.1-t2v-14b
    StatusRunner parity pending

    text-to-video

    Runtime profilewan2.1-t2v-14b

    Pipeline bindingwan2.1-t2v-14b

  • wan2.1-i2v-14b-480p
    StatusRunner parity pending

    image-to-video

    Runtime profilewan2.1-i2v-14b-480p

    Pipeline bindingwan2.1-i2v-14b-480p

  • wan2.1-i2v-14b-720p
    StatusRunner parity pending

    image-to-video

    Runtime profilewan2.1-i2v-14b-720p

    Pipeline bindingwan2.1-i2v-14b-720p

What it is for

Start on wan2.1-t2v-1.3b. Launch worldfoundry-eval run wan2.1-t2v-1.3b. Use wan2.1-t2v-14b or I2V 480p/720p only when you accept Runner parity pending on those lines.

Iterate prompts on the 1.3B line when you need a cheap, verified artifact. Wan2.1 is the default backbone when a later catalog entry (VACE, Helios, camera-control world models) says it is Wan-derived. Compare clip quality on VBench and physical plausibility on Physics-IQ — those are the suites Wan releases are commonly discussed against, not ingested WorldFoundry leaderboards.

Runtime contract

Required prompt; optional image / video; actions is seed. Verified 1.3B T2V: 832×480, 81 frames, 50 steps, guidance 6.0, 16 fps, shift 8.0 through Wan2p1T2VPipeline (wan2.1-t2v-1.3b-832x480-81f). Artifact: wan2.1-t2v-1.3b-832x480-81f.mp4. 14B T2V default 1280×720, shift 5.0; I2V 480p 832×480 / 81 frames / 40 steps / guidance 5.0 / shift 3.0; I2V 720p 1280×720, shift 5.0. Catalog tasks are text-to-video and image-to-video only.

Wan-VAE encodes 1080p of arbitrary length so the DiT denoises in latent space. Text is UMT5. I2V needs the image-encoder checkpoint role.

Running in WorldFoundry

Unified worldfoundry-unified-cu128 (Python 3.11, CUDA 12.8). Stage Wan-AI checkpoints locally. License Apache-2.0. Catalog VRAM 8 GB on the 1.3B line. Runtime integration is recorded for the family. A prior local 81-frame T2V 1.3B artifact is recorded with verified demo and runner parity; 14B T2V and both I2V lines still lack output parity. No public parity demo script is packaged, and reproducing the 1.3B artifact requires the staged checkpoint and runtime environment.

Generate the 1.3B T2V clip, then score the folder on VBench custom_input / aesthetic_quality — the only VBench GPU dimension this repo verified. Read aesthetic_quality in tmp/vbench/wan2.1/eval/scorecard.json.