Kandinsky 6.0 VSR

Kandinsky 6.0 VSR — Video Super-Resolution (Video-to-Video)

What is Kandinsky 6.0 VSR?

Kandinsky 6.0 VSR is a prompt-free video super-resolution model that upscales and restores short clips, recovering fine detail and textures while keeping frames temporally consistent. Part of Sber's open Kandinsky 6.0 Video family, it takes a soft, low-resolution clip — up to 121 frames (about five seconds at 24 fps) — and rebuilds it at 2x, 2.25x or 4x, so a 480p source becomes a clean 1080p result. It encodes the source, upscales in latent space, denoises overlapping tiles and blends them back together, which is why recovered detail stays stable frame to frame instead of shimmering. It works on both AI-generated drafts and real camera footage.

Key Features

• •Fine-detail and texture recovery: feather barbs, fabric weave, foliage, skin and other high-frequency detail that turns mushy at low resolution resolves cleanly after upscaling.
• •Flicker-free temporal consistency: detail is reconstructed across overlapping tiles so moving subjects stay stable, with no boiling or shimmer.
• •2x, 2.25x and 4x upscaling: 2.25x handles any input size (480p to 1080p); 2x and 4x need width and height that are multiples of 16.
• •Pro and Lite modes: Pro (5 steps) restores the most detail; Lite (2 steps) is faster and cheaper with comparable results.

Best Use Cases

Kandinsky 6.0 VSR is built to be a finishing stage. Use it to take AI-generated video drafts — including Kandinsky 6.0 Video Lite and Pro output, which generates at SD resolution — up to Full HD. It also restores and sharpens short real-world clips: wildlife and nature macro, product shots, textiles and craft close-ups, archival footage, and social clips that need to read crisply at 1080p. In testing, a soft 848x480 wildlife clip upscaled to 1920x1088 in Pro mode rebuilt feather and beak texture without flicker across the motion, and Lite mode produced a near-identical restore at lower cost.

Input Tips and Output Quality

Trim sources to 121 frames or fewer before sending; longer clips are rejected. Choose 2.25x for any-size 480p-to-1080p work; switch to 2x or 4x only when your dimensions are multiples of 16. Cleaner sources upscale better — heavy compression and noise limit what any VSR can recover. Start in Lite for drafts and batches, move to Pro for finals. Because the model restores rather than invents, framing and motion from the source carry through unchanged.

How to Use Kandinsky 6.0 VSR

Kandinsky 6.0 VSR upscales and restores short video clips without a prompt. You give it a source video, pick a quality mode and an upscale factor, and it returns a sharper, higher-resolution clip with recovered texture and stable, flicker-free frames. Everything hinges on three inputs: the source clip, mode, and upscale_factor.

How it works

The model encodes your clip, upscales it in latent space, then denoises overlapping tiles and blends them, so detail is rebuilt consistently across frames. It handles up to 121 frames (about five seconds at 24 fps). Trim anything longer before you send it, or the request is rejected.

Recommended Settings

• •mode — pro (5 steps) for maximum detail on finals; lite (2 steps) for faster, cheaper drafts and batch jobs. On tested clips Lite stayed close to Pro.
• •upscale_factor — 2.25 is the default and the most flexible: it takes any-size 480p to 1080p. Use 2 or 4 only when your input width and height are multiples of 16.
• •num_inference_steps — leave empty to use the mode default (pro 5, lite 2); raise it only if texture looks under-recovered.
• •seed — keep -1 for variety, or set an integer to reproduce a run.

For a 480p source, start with mode: pro, upscale_factor: 2.25 to reach 1080p. Expect roughly three to five minutes per five-second clip; 4x is a heavier render.

Example upscales to try

Each example below ran end to end at the settings shown, on the same soft 848x480 wildlife clip, so you can compare tiers on one source — motion and framing carried over unchanged.

• •Detail recovery (default). Settings: mode: pro, upscale_factor: 2.25. The 480p clip reached 1920x1088 with feather barbs, facial lines and beak keratin rebuilt and no flicker across the head-turn. Tip: this is the go-to for any-size 480p sources.
• •Lite mode. Settings: mode: lite, upscale_factor: 2.25. Restored nearly as much detail as Pro at lower cost. Tip: use Lite first for drafts and batches, and switch to Pro only if texture looks soft.
• •4x upscale. Settings: mode: pro, upscale_factor: 4. Took 848x480 to 3392x1920 with crisp micro-texture and no artifacts. Tip: 2x and 4x need input dimensions that are multiples of 16, and 4x is a heavier, slower render — reserve it for finals.

Beyond these, the model is built to "finish a Kandinsky 6.0 Video draft at Full HD", "restore a short real-world clip to 1080p", or "sharpen a compressed social clip" — supply a clean source and the recovered detail holds up.

Does Kandinsky 6.0 VSR need a prompt?

No. It is a prompt-free upscaler — you supply only a video and the mode and factor.

What is the maximum length?

About five seconds, or 121 frames at 24 fps. Trim longer footage first.

Which upscale factor should I use?

2.25x for any-size 480p to 1080p; 2x and 4x require input dimensions that are multiples of 16.

Does it work on real footage or only AI video?

Both. It restores camera footage and AI-generated drafts alike.