Kandinsky 5.0 Pro: 19B Open Video Model and 8 Prompts

Most image-to-video results fail in the same quiet way. The first frame is your photo. By frame forty it's a cousin of your photo: the face has softened, the fabric has lost its weave, and the camera has wandered off somewhere you never asked it to go. The motion looks fine. The picture you started with is gone.

Kandinsky 5.0 Pro is the big video model from Kandinsky Lab at Sber AI, and it's built for exactly that job: take one still, keep it recognisable, and move it with intent. It's also open. The weights sit on GitHub and Hugging Face under an MIT license, which is rare for a model in its weight class. Search interest in "kandinsky 5.0 pro" has jumped in the last week, so here's what the model actually is, what the image-to-video version on each::labs accepts, and eight prompts you can paste in today.

What Kandinsky 5.0 Pro Actually Is

Kandinsky 5.0 is a family, not a single model, and the naming matters. The technical report (Kandinsky 5.0: A Family of Foundation Models for Image and Video Generation, submitted to arXiv in November 2025) describes three lines. Image Lite is a 6B parameter image generator. Video Lite is a 2B parameter model for text-to-video and image-to-video, tuned for speed. Video Pro is the heavyweight: a line-up of 19B parameter models aimed at the highest video quality the family can deliver.

Pro is what you want when the clip has to hold up on a real screen. Kandinsky Lab describes it as generating HD video from English and Russian prompts with controllable camera motion, and it ships in two flavours: an SFT checkpoint for the best output and a Pretrain checkpoint meant for teams that want to fine-tune. The Pro text-to-video and image-to-video models were open-sourced together on November 20, 2025. A few weeks later, on December 12, the team reported Video Pro as the top open-source text-to-video model on LMArena.

Under the hood it's a latent diffusion model trained with flow matching, built on a diffusion transformer that cross-attends to text embeddings from Qwen2.5-VL and CLIP, with the HunyuanVideo 3D VAE handling video encoding and decoding. The training recipe runs from large-scale pretraining through supervised fine-tuning to reinforcement-learning-based post-training. You don't need any of that to write a prompt. It does explain why Pro tends to read a scene description closely instead of skimming it for nouns.

Pro vs Lite: Pick the Right Kandinsky

The gap between the two video lines isn't subtle. Lite is 2B parameters, Pro is 19B, close to ten times the capacity. Lite is the one the team distilled for speed (its diffusion-distilled variant is listed as six times faster), and it's the one they pitch as the open model with the best grasp of Russian cultural concepts. Pro is the one they put up against Veo 3, Veo 3 fast and Wan 2.2 A14B in side-by-side human evaluations, in both text-to-video and image-to-video.

Length is the other spec people ask about. The family generates 5-second and 10-second clips, and the 10-second versions use a sparse attention scheme called NABLA to keep long generations tractable. The released Pro image-to-video checkpoints are the 5-second ones, in HD and SD. So think of Kandinsky 5.0 Pro image-to-video as a short-shot tool: each render gives you one strong beat, and you cut the beats together into something longer.

My take: if you're animating a single hero image for a product page, a pitch or a social post, Lite is a false economy. The extra capacity in Pro shows up exactly where image-to-video usually breaks, in faces, textures and the small physical details that tell a viewer the footage is real.

Kandinsky 5.0 Pro on each::labs: The Real Parameters

On each::labs you run Kandinsky 5 Pro Image to Video, the Pro image-to-video model, from the Kandinsky 5 family page. The request is refreshingly small. Two fields are required: prompt, the motion and scene description, and image_url, the still you want animated. Everything else has a sensible default.

resolution takes 512P (the default) or 1024P. num_inference_steps defaults to 28 denoising steps. acceleration is either regular (the default) or none. Output is an MP4, and the model page lists 16:9, 9:16 and square framing, with an average processing time of about three minutes.

There's no duration field and no negative prompt, and that's a design choice worth respecting. With fewer knobs, the prompt and the source image do almost all of the steering. A minimal call looks like this:

curl -X POST https://api.eachlabs.ai/v1/prediction/ \
  -H "Authorization: Bearer $EACHLABS_API_KEY" \
  -H "Content-Type: application/json" \
  --data '{
    "model": "kandinsky5-pro-image-to-video",
    "input": {
      "prompt": "The woman turns her head toward the window and smiles slightly. Soft daylight, slow push-in, steady camera.",
      "image_url": "https://example.com/portrait.jpg",
      "resolution": "1024P",
      "num_inference_steps": 28,
      "acceleration": "regular"
    }
  }'

The response hands back a prediction ID, which you poll or receive by webhook. Build for asynchronous results from day one and you'll never have to retrofit it.

How to Prompt Kandinsky 5.0 Pro for Image to Video

The single most useful rule: don't describe the photo. The model can already see it. Every word you spend on "a woman in a red coat on a bridge" is a word you didn't spend on what she does next. Describe the change, not the scene.

Write in this order. First the subject's action, in one clear verb. Then the secondary motion, the things that move because the world moves: hair, steam, leaves, water, fabric. Then the camera, named plainly (slow push-in, static shot, gentle pan left, orbit). Last, the light and mood, only if you want them to shift. Kandinsky 5.0 Pro accepts English and Russian, so write in whichever language you think in. Keep it to two or three sentences. Five seconds is short, and one well-defined action reads better than three crammed together.

Name the camera even when you want none. "Static shot, locked-off camera" is an instruction. Leaving the camera out is an invitation to drift. We covered the same principle across models in our guide to writing the motion in image-to-video prompts, and it applies doubly to a model this attentive to text.

8 Copyable Kandinsky 5.0 Pro Prompts

Each of these assumes you've uploaded a matching source image. Swap the nouns, keep the structure.

1. Portrait, subtle life. For headshots and character art that need to breathe without becoming a different person.

She blinks slowly and lifts her gaze to the camera, a faint smile forming. A loose strand of hair moves in a light breeze. Static shot, soft natural window light.

2. Product hero. For a packshot on a clean background.

The bottle rotates slowly a quarter turn to the right as a soft highlight travels across the glass. Condensation droplets glisten. Slow orbit around the product, studio lighting, clean background stays still.

3. Landscape loop. For hero banners and ambient backgrounds.

Clouds drift across the peaks and their shadows slide over the valley. Tall grass in the foreground sways gently. Locked-off camera, late afternoon golden light.

4. Food, steam and pour. For menus and recipe content.

Steam rises in thin curls from the bowl as a spoon slowly lifts noodles out of the broth. Slow push-in toward the bowl, warm overhead light, shallow depth of field.

5. Illustration brought to life. For painted or drawn artwork, where the style must survive the motion.

The lanterns sway softly and their glow flickers. Snow falls gently across the street while the painted texture stays intact. Very slow pan to the right.

6. Fashion, fabric in motion. For lookbook stills.

The model takes one step forward and turns her shoulder toward the camera. The silk dress ripples and catches the light. Slow dolly back, overcast daylight, city street behind her stays sharp.

7. Architecture, reveal. For interiors and real-estate renders.

Sunlight slowly sweeps across the wooden floor as sheer curtains billow inward. Smooth forward glide through the room toward the window, steady camera.

8. Vintage machine, mechanical motion. For heritage brands and editorial work.

The steam locomotive rolls forward along the curved track as thick smoke drifts back over the trees. Slow cinematic push forward, steady camera, soft morning haze over the valley.

For more patterns that carry over between models, these AI video prompt examples are a good next stop.

Dial In Resolution, Steps and Speed

Treat 512P as your sketchpad. Run a new prompt there first, check that the action and the camera are what you meant, and only then render the keeper at 1024P. Most bad generations are prompt problems, not resolution problems, and you'll find them faster at the smaller size.

Leave steps at 28 until you have a reason to move. The open-source Pro checkpoints are documented at 100 function evaluations, so the each::labs default of 28 is the quicker route by a wide margin. If a render shows smeared fine detail on a face or a fabric, raise the step count and set acceleration to none on that one final pass. Don't do it for drafts. You'll only make the iteration loop slower without learning anything new.

Pick your source image like you'd pick a first frame on a shoot. Sharp subject, clear separation from the background, light that already suggests a direction. The model page notes that subject consistency (facial features, object textures) is a strength, and it can only preserve what's actually there.

When to Pick Kandinsky 5.0 Pro vs Other Image to Video Models

The image-to-video catalog on each::labs is deep, and Kandinsky 5.0 Pro earns its slot for specific reasons rather than general ones.

Reach for it when the source image has to survive. Portraits, product shots and illustrated art all live or die on detail retention, and a 19B model with a short, focused clip length puts its capacity into keeping your frame intact. Reach for it when openness matters to your roadmap: MIT-licensed weights mean the model you prototype against is the same model you could later fine-tune or self-host from the Pretrain checkpoint. And reach for it when your prompts come in Russian as well as English.

Look elsewhere when the brief is different. Kandinsky Lab chose Wan 2.2 A14B Image to Video as one of its side-by-side baselines, and it's the natural open-weight model to run on the same photo when you want a second opinion. When you'd rather see how a closed flagship reads the same frame, Kling v3 Pro Image to Video is the one to test next. Because everything runs behind the same API, comparing them is a change of model string, and you can chain the winner into a larger pipeline with each::labs flows.

Frequently Asked Questions

Is Kandinsky 5.0 Pro open source?

Yes, under MIT. Kandinsky Lab published the Video Pro text-to-video and image-to-video checkpoints, in SFT and Pretrain variants, along with the inference code on GitHub and Hugging Face in November 2025, and Video Pro was added to the diffusers library that December. The official Kandinsky 5 repository has the full list of checkpoints.

How long are Kandinsky 5.0 Pro videos?

Short by design. The Kandinsky 5.0 video family offers 5-second and 10-second models, and the released Pro image-to-video checkpoints are the 5-second versions. Plan one action per clip and assemble longer sequences in the edit.

What resolution does Kandinsky 5.0 Pro image to video output on each::labs?

Two settings: 512P by default and 1024P for finals. The model page lists 16:9, 9:16 and square framing, so the shape of your source image is the first composition decision you make.

What's the difference between Kandinsky 5.0 Pro and Lite?

Mostly capacity and intent. Video Pro is a 19B parameter line built for the highest quality output, while Video Lite is a 2B parameter model with distilled variants for speed. For animating a single image that has to stay faithful, Pro is the one on each::labs today.