VideoCrafter
Tencent AI Lab VideoCrafter family page. Default is VideoCrafter2 T2V at 512×320 / 16 frames / 28 fps / 50 DDIM steps; VC1 T2V and I2V are separate pending routes.
cogvideo is not in this catalog (the later CogVideoX card is a different family). This page is the unused VideoCrafter family entry — Tencent AI Lab's U-Net latent video toolbox. It aggregates three in-tree infer-only runtimes: VideoCrafter2 T2V (the default launch), VideoCrafter1 T2V, and VideoCrafter1 I2V. Checkpoint provenance lives here so the sibling pages do not duplicate Hugging Face aliases. Training, dataset, and download helpers are not packaged.
No official teaser or method figure is checked into this docs tree.
The paper field is VideoCrafter2: Overcoming Data Limitations for High-Quality Video Diffusion Models. Project: ailab-cvc.github.io/videocrafter2. Code: VideoCrafter/VideoCrafter. paper.png is the title thumb only.
Build your run
VideoCrafter2T2VPipelineworldfoundry.pipelines.videocrafter.pipeline_videocrafter2_t2vdefaultworldfoundry.pipelineworldfoundry-eval run \
videocrafter \
--pipeline.task-profile default \
--jsonWhat it is for
Start on VideoCrafter2 T2V (VideoCrafter2T2VPipeline): required prompt, artifact videocrafter2-t2v.mp4. That is the license-cleanest card (Apache-2.0) and the family default (worldfoundry-eval run videocrafter).
Use videocrafter1-t2v for the older 1024×576 text-to-video weights, and videocrafter1-i2v to animate a still. There is no V2V route. Compare the companion suite EvalCrafter and general T2V quality on VBench — listed pointers, not ingested ranks.
Inference
Sampler is DDIM (ddim_steps=50, ddim_eta=1.0, unconditional_guidance_scale=12.0 unless you override). Height and width must be divisible by 16. Frame count defaults to the U-Net temporal_length of 16.
| Route | Task | Size / frames / fps | Checkpoint | Parity |
|---|---|---|---|---|
videocrafter2-t2v (default) | T2V | 512×320, 16f, 28 fps | VideoCrafter/VideoCrafter2 | pending |
videocrafter1-t2v | T2V | 1024×576, 16f, 8 fps | VideoCrafter/Text2Video-1024 | pending |
videocrafter1-i2v | I2V | 512×320, 16f, 8 fps | VideoCrafter/Image2Video-512 | pending |
Catalog VRAM is unset. The family I2V task still names videocrafter2-t2v.mp4 as the recorded artifact — that is a catalog quirk; I2V actually goes through the videocrafter1-i2v pipeline.
Running in WorldFoundry
Unified env worldfoundry-unified-cu128 (Python 3.11, CUDA 12.8). Stage the matching model.ckpt under $WORLDFOUNDRY_HFD_ROOT (VideoCrafter--VideoCrafter2, VideoCrafter--Text2Video-1024, or VideoCrafter--Image2Video-512). Older Text2Video / Image2Video cards expose no license in cardData — do not infer Apache-2.0 from the org name. Status: Runner parity pending.
