How to Use Kandinsky 5 to Generate Video and Images Locally

Kandinsky 5 is an open-weight generative model that produces short video clips and still images from text prompts, and it runs on a single consumer GPU. It’s released by Sber AI and covers both text-to-video and text-to-image in one family.

Open video models used to be a joke — blurry, five frames, unusable. That changed. Kandinsky 5 is part of the new wave: weights you can download, quality that’s genuinely watchable for short clips, and hardware needs that don’t require a data center.

What Can Kandinsky 5 Actually Do?

Two modes ship in the family. The video model generates clips of a few seconds at 720p-class resolution. The image model outputs 1024×1024 stills. Both take natural-language prompts and support negative prompts.

It’s tuned for stylized and cinematic looks — good at landscapes, motion, and moody lighting, less reliable at crisp text and precise anatomy.

What Hardware Do You Need?

• Minimum: a 12GB GPU for the image model; 16-24GB for video.
• Comfortable: 24GB VRAM gives you longer clips and higher resolution without quantization headaches.
• Speed: a short 5-second clip takes roughly 2-4 minutes on a 4090 at default settings.

Apple Silicon works via MLX ports if you’re patient. Expect it to be slower than an NVIDIA card, but it runs.

How Do You Run Your First Generation?

1. Grab the weights from Hugging Face and set up ComfyUI — there are ready-made Kandinsky nodes.
2. Load the video or image checkpoint depending on your target output.
3. Write a prompt with a shot type: “wide aerial shot of a foggy forest, slow dolly forward.” Shot and camera language matters here.
4. Add a negative prompt for common artifacts: “warping, extra fingers, text.”
5. Generate, then upscale — native output is softer than a final render should be.

Is Kandinsky 5 Free?

Yes to download and use. The weights are openly licensed, including commercial use, though you should check the current license terms since they’ve been revised between versions. There’s no per-generation fee — your only cost is electricity or GPU rental.

How Does It Compare to Closed Video Models?

Honestly? Veo and Sora still win on prompt adherence and realism. Kandinsky 5 closes the gap on short stylized clips and wins outright on the two things open models always win: cost at volume and the ability to fine-tune on your own footage.

Pick it if you’re building a content pipeline, want to train on a specific aesthetic, or can’t send prompts to a cloud API. Skip it if you need photoreal humans with correct hands on the first try.

Quick Summary

• Kandinsky 5 generates video clips and 1024px images from text.
• Runs on a single 12-24GB GPU; a 5-second clip takes 2-4 minutes on a 4090.
• Open weights with commercial use; no per-generation cost.
• Great for stylized short clips, weaker than Veo/Sora on realism and text.