WorldFoundry
English中文

Kandinsky 6.0 pro-distill

Kandinsky 6.0 pro-distill from Kandinsky6: text-to-video and image-to-video; Runner parity pending; dedicated `worldfoundry-kandinsky6-cu126`.

Edit this pageGitHubMarkdown

Kandinsky 6.0 Pro-distill generates synchronized video and 44.1 kHz audio from text, with optional first-frame image conditioning. WorldFoundry runs the Pro transformer with 10 PiFlow steps through its native diffusion components.

No official teaser or method figure is checked into this docs tree.

Code: kandinskylab/kandinsky-6.

Build your run

Pipeline
Kandinsky6Pipelineworldfoundry.pipelines.kandinsky6.pipeline_kandinsky6
TaskDefault default
Environmentworldfoundry-kandinsky6-cu126
DeviceCUDA 12.6
Runnerworldfoundry.pipeline
Loadingfrom_pretrained
Invocationnative_or_standard_invocation
worldfoundry-eval run \
  kandinsky6-pro-distill \
  --pipeline.task-profile default \
  --json

What it is for

Use kandinsky6-pro-distill for the family's joint T2V/I2V inference path. The catalog records both task profiles on this page.

Catalog tasks include text-to-video, image-to-video, and audio-video-generation. Catalog aliases include kandinsky6 and kandinsky-6. Documented modalities: text and image in, video and audio out.

Runtime contract

Optional prompt, input_path, interactions, operator_kwargs, and ref_image_path; artifacts generated_video and kandinsky6-pro-distill.mp4.

CLI task profiles: default. WorldFoundry binds Kandinsky6Pipeline.

Family status: Runner parity pending.

Text-to-video passed on one H100 80GB with SDPA, component offload, and seed 42: 121 frames at 864x480 and 24 fps, with muxed audio. Weight loading took 85.2 seconds and generation took 140.4 seconds in the pinned Python 3.10 / CUDA 12.6 environment. The validation report is outputs/kandinsky6-pro-distill-validation/cuda-validation.json.

Default text-to-video generation passed on one H100 80GB with SDPA and component offload.

Running in WorldFoundry

Launch worldfoundry-eval run kandinsky6-pro-distill --pipeline.task-profile default 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 dedicated environment worldfoundry-kandinsky6-cu126 (Python 3.10, CUDA 12.6, torch==2.9.0+cu126).

Recorded weight repositories include kandinskylab/Kandinsky-6.0-Pro-distill-5s-Diffusers. Prepare the environment and check local assets with bash scripts/inference/prepare_model_infer.sh kandinsky6-pro-distill. Add --download to fetch missing public checkpoints.

Status: Runner parity pending.