Kandinsky 6.0 Video: GPU Requirements, Speed and Cost

Released on October 6, 2026, the family combines video and 44 kHz audio generation, including lip sync. A reference image can guide the opening frame. Full HD output uses a separate video super-resolution stage, so include that stage when planning the finished workflow.

Lite and Pro describe the model sizes. Distilled identifies a faster sampling version of either size. For a first run, choose a generation checkpoint rather than selecting a pretraining checkpoint simply because its file is available.

The official repository includes an RTX 5060 Ti 16GB preset for Pro-distill. It uses block offloading, Sage attention and a quantized Qwen text embedder. Blocks move between system memory and the GPU as needed; the full pipeline does not stay in 16GB of VRAM.

Use that specific preset as the starting point on the matching card. System RAM, storage, the text encoder and super-resolution also need space. The preset does not specify a system-RAM minimum, so check peak memory during a complete run before sizing a machine or a batch.

For another supported NVIDIA card, select its own device configuration from the repository. Match the attention backend, runtime and offload configuration as well as VRAM capacity.

Kandinsky Lab reports these times for the non-distilled models after warmup, excluding weight loading and MP4 encoding. The consumer cards use block offloading; the H100 uses module offloading. These measurements describe a different checkpoint from the Pro-distill preset above.

The standalone package requires Python 3.13 or 3.14, an NVIDIA GPU, uv and just. Its setup selects the GPU dependencies. On Ampere, Ada and consumer Blackwell cards, SageAttention is compiled, so the CUDA compiler nvcc must be available on PATH.

The command uses the Pro-distill preset’s default 864×480 output, 121 frames, 10 steps and CFG 1, without super-resolution. Add the separate super-resolution stage for Full HD. Its settings differ from the non-distilled HD benchmark above.

The example below chooses the RTX 5060 Ti configuration explicitly. Replace the configuration with the preset for your GPU. The download command fetches Pro-distill; the generation command also downloads any checkpoint selected by the configuration if it is missing.

For a visual workflow, the official extension requires ComfyUI 0.38.0 or later and Python 3.10 or later. In ComfyUI Manager, look for kandinsky6 and kandinsky6-sr from kandinskylab, install both and restart. If either is absent, follow the official guide’s manual-install route. ComfyUI handles model loading and offloading.

Open Workflow → Browse Templates → kandinsky6, then choose the text-to-video-and-audio or image-to-video-and-audio template. Use Download models in the workflow note before the first run. Installation alone does not download the weights. For an image-led shot, replace the example portrait in Load Image.

The bundled Pro-distill workflow uses PiFlow with 10 steps and CFG 1. Edit the video and audio captions in Beautify Prompt. Place exact speech between <S> and <E> in the video caption, and describe the voice in the audio caption. Disable beautification when you want to test the captions exactly as written.

Kandinsky Lab added vLLM-Omni support on October 6. Use a separate Python 3.12 environment with vLLM 0.31.0 and the current vLLM-Omni source, following the repository’s installation guide. Keep it separate from the standalone just setup environment above.

The documented checkpoint is Kandinsky-6.0-Pro-5s-Diffusers, the non-distilled Pro model. Its example uses 864×480, 125 frames at 24 fps, 50 steps and guidance 5 with audio. Enable CPU offloading even on an 80GB GPU, as the guide requires.

Choose the offline examples for a batch or the online server for a service using /v1/videos. Use the documented image-to-video example for image input; its masked-tail-frame handling differs from simply passing an ordinary first frame. Measure this backend separately before applying the earlier runtime or cost examples to it.

Keep the first shot to one action that can finish within five seconds. For a product still, ask for a gentle camera move and one physical change. For dialogue, use one short line. Listen for the requested sound, watch the contact between objects, and inspect the face through the final frame.

A practical image-to-video prompt is: “The small desk fan in the reference image starts spinning. Its base stays fixed on the desk while the blades accelerate. A locked close-up with warm window light. Audio: a switch click followed by a steady fan hum.” This makes motion, object stability and sound timing easy to judge.

Keep the prompt, input image, seed, checkpoint and settings with the accepted result. Compare Lite and Pro with the same brief, then compare distilled and non-distilled versions separately. If the action fails, simplify it before paying for another upscale.

Self-hosting has a compute cost rather than a fixed Kandinsky charge per video. For a rental, multiply the billed machine hours by its hourly price. Include setup, loading, unused rental time and every attempt, then divide the batch bill by the number of clips you keep.

As a hypothetical example, at $1.00 per GPU-hour, the published non-distilled RTX 5090 HD timings correspond to about $0.08 for Lite or $0.18 for Pro per five-second generation. This assumes the reported warm runtime and excludes loading, encoding, storage and other charges.

Three Pro attempts at that same assumed rate and runtime use about $0.53 of generation time for one chosen shot. Add any separately run upscale and the time required to review and edit it. On a GPU you own, measure whole-system energy use and account for the machine’s purchase cost and competing workloads.

The Kandinsky 6.0 repository uses the MIT license, and the official Lite and Pro-distill model cards also list MIT. It permits commercial use, modification and redistribution subject to preserving the required copyright and permission notices.

A complete workflow also includes text encoders and other components. Keep their accompanying license notices with the files you deploy or redistribute. For client work, use reference images and voices you have permission to use.

Kandinsky offers MIT-licensed model releases, a five-second workflow and published timings for several NVIDIA GPUs. It is a useful option when you want control over a short-shot pipeline and can accommodate its generation time.

LTX 2.5 offers a hosted API with published per-second prices as well as downloadable weights. It is worth comparing when you need longer clips, a hosted production path or multi-shot generation. Its model weights use a Community License with a revenue threshold, so include licensing alongside image quality and compute cost in your choice.

For either model, compare the total cost of a usable shot with your required sound and resolution. An inexpensive generation that needs several rerolls can cost more than a higher-priced first take you can use.

Can I try it before setting up a GPU? Kandinsky Lab provides a Pro-distill demo on Hugging Face Spaces. It is a convenient first look at the model’s output before downloading the local workflow.

Does a five-second output mean five seconds of processing? No. Clip duration and generation time are different quantities; the benchmark table above shows processing times in minutes.