Skip to main content
AI tutorials

Kandinsky 6.0 Tutorial: Generate Synchronized Video and Audio with AI

Learn how to install Kandinsky 6.0, download a model checkpoint, generate synchronized video and 44 kHz audio from a text prompt, select GPU presets, use ComfyUI, and optimize your workflow with the project’s performance options.

About 14 min read
Kandinsky 6.0 Tutorial: Generate Synchronized Video and Audio with AI

Kandinsky 6.0 is a family of diffusion models for synchronized text-to-audio-video generation. Kandinsky 6.0 Video Lite has 3 billion parameters, while Kandinsky 6.0 Video Pro has 29 billion parameters. Both models can create five-second clips with synchronized 44 kHz audio, including lip-sync, in text-to-audio-video and image-to-audio-video modes.

A plugin super-resolution model can increase the output resolution to Full HD, or 1920×1080. In this tutorial, you will install the project, download a checkpoint, generate your first clip, choose a device preset, and review additional ways to run Kandinsky 6.0.

Kandinsky 6.0 project preview

What Kandinsky 6.0 provides

Kandinsky 6.0 is designed for synchronized multimodal generation rather than silent video synthesis alone. Its models generate video and audio together, supporting both text-driven and image-driven workflows.

  • Video and audio generation: Create five-second clips with synchronized 44 kHz audio.
  • Lip-sync: Generated content can include synchronized lip movements.
  • Two model sizes: Choose between the 3B-parameter Video Lite family and the 29B-parameter Video Pro family.
  • Two input modes: Use text-to-audio-video or image-to-audio-video generation.
  • Super-resolution: Use the plugin-in super-resolution model to raise output resolution to Full HD.

Requirements

The local quick-start workflow requires an NVIDIA GPU and Python 3.13 or 3.14. You will also need uv and just.

The setup command detects the GPU and installs an appropriate PyTorch build. Hopper GPUs, including H100 and H200 devices, use PyTorch for CUDA 13.0 with FlashAttention 3. Ampere, Ada, and consumer Blackwell GPUs use PyTorch for CUDA 13.0 and SageAttention 2.2.0.

If you are using an Ampere, Ada, or consumer Blackwell GPU, nvcc must be available on your PATH. The setup expects CUDA_HOME to default to /usr/local/cuda.

Install Kandinsky 6.0

1. Clone the repository

Clone the repository and move into the project directory:

git clone https://github.com/kandinskylab/kandinsky-6.git
cd k6_video
just setup

The just setup command reads your GPU and installs one of the supported PyTorch and attention configurations.

2. Download a model

Download the Pro distilled checkpoint with:

just download pro-distill

On the first run, the project downloads Kandinsky-6.0-Pro-distill-5s from Hugging Face. By default, model weights are stored under ~/.cache/kandinsky/weights.

You can change this base location with the KANDINSKY_HOME environment variable. When it is set, weights are stored in $KANDINSKY_HOME/weights.

Generate your first video

After setup and model download, generate a clip from a text prompt:

just generate "a cat on a mat"

The resulting MP4 is written to a timestamped output directory:

$KANDINSKY_HOME/outputs/generate_<YYYY-MM-DDTHH-MM-SS>/generations/output.mp4

If KANDINSKY_HOME is not set, the project uses ~/.cache/kandinsky as its default home directory. The same generation directory also contains the expanded prompt in expanded_prompt.txt, the run configuration in launch.json, and the logs/ and profiles/ directories.

The generate command downloads the checkpoint named in the configuration if it is not already available locally.

Choose a GPU preset

If your GPU needs a different configuration, pass a device preset with --config. For example:

just generate "a cat on a mat" --config kandinsky/configs/devices/rtx-5090.yaml --out clip.mp4

Device presets are stored in kandinsky/configs/devices/. Select the preset that matches your hardware and workload.

Available model catalog names

The quick-start documentation lists these model catalog names:

  • pro
  • pro-pretrain
  • lite
  • lite-distill
  • lite-pretrain

The default first-run example uses pro-distill, which downloads the Pro distilled five-second checkpoint.

Use Kandinsky 6.0 with ComfyUI

Kandinsky 6.0 also supports ComfyUI. Install kandinsky6 and kandinsky6-sr through ComfyUI Manager, then restart ComfyUI.

The repository’s comfyui/ directory contains the extension source code. Manual copying is not required. For the project-specific installation details, consult the ComfyUI setup guide.

Performance and optimization tips

Generation time depends heavily on the model family, resolution, GPU, and offloading strategy. The project’s benchmarks measure the working time for a five-second clip after warmup; weight loading and MP4 encoding are excluded.

The documented hardware configurations use module offload on RTX PRO 6000, A100 80 GB, and H100 80 GB systems. Consumer cards use block offload.

Attention implementation also varies by GPU. RTX 4090, RTX 5060 Ti, RTX 5080, RTX 5090, RTX PRO 6000, and A100 80 GB systems use SageAttention2++ for visual self-attention. H100 systems use FlashAttention 3.

Interpret benchmark results

The reported times are in seconds for a five-second clip using the non-distilled model. The listed configurations include Lite SD, Lite HD, Lite Full HD, Pro SD, Pro HD, and Pro Full HD. In general, Lite configurations require less time than Pro configurations, while Full HD generation requires more time than SD or HD generation on the same hardware.

For the fastest local experiments, begin with a Lite or distilled checkpoint and an appropriate device preset. Use Pro or Full HD configurations when the additional model capacity or resolution is important for your result.

Practical workflow

  1. Confirm that your system has an NVIDIA GPU and Python 3.13 or 3.14.
  2. Install uv and just.
  3. Clone the repository and run just setup.
  4. Download a checkpoint with just download pro-distill or another supported catalog name.
  5. Run just generate with a descriptive prompt.
  6. Locate the generated MP4 in the timestamped output directory.
  7. Review expanded_prompt.txt, launch.json, logs, and profiles when you need to inspect a run.
  8. Switch to a device preset or ComfyUI when your workflow requires different hardware settings or node-based control.

Conclusion

Kandinsky 6.0 provides local text-to-audio-video and image-to-audio-video generation with synchronized audio, lip-sync support, and optional super-resolution. The simplest path is to install the dependencies with just setup, download a checkpoint, and run a prompt with just generate. For more control, use the device presets, inspect the generated run files, or install the Kandinsky 6.0 extensions through ComfyUI Manager.

For more information, visit the KandinskyLab website, the Diffusers documentation, or the Hugging Face demo.