SeedVR2-3B — Video Upscaling AI Model
SeedVR2 3B lifts video to anywhere from 720p to 4K, holding detail steady from frame to frame instead of shimmering. The lighter of the pair — two to five minutes, and longer once tiling kicks in at 4K. Input is capped at 495 frames — about 16 seconds at 30fps, 20 at 24fps — so shorten a longer clip before you upload it.
Demo
Enlarged to 720x1264 by ordinary resizing — 1:1 crop
Enlarged by SeedVR2 — the same crop, same pixel size
What it does
SeedVR2-3B is a video restoration and enhancement model developed by ByteDance Seed. It takes damaged, blurry, compressed, or low-resolution video footage and restores it to sharp, high-quality output.
What sets SeedVR2 apart from previous AI video restoration tools is its speed: where most comparable tools require many processing passes to produce a good result, SeedVR2 achieves the same or better quality in a single pass. This makes it significantly faster and more cost-effective to run at scale, without meaningful sacrifice in output quality.
When to use it
- Old or degraded footage — restores heavily compressed, blurry, noisy or low-resolution video to sharp output
- Slow restoration pipelines — completes restoration in a single step instead of many, cutting processing time
- High-resolution output — scales up without the visual inconsistencies other tools commonly show
- Real-world content — designed for real footage, not just clean lab test videos
Details
- CategoryVideo Upscaling
- Year2025
- LicenseApache-2.0
Compliance & Provenance
- ProviderOpen-source (ByteDance) · Specialized
- EU AI Act RiskMinimal Risk
- Art. 50 TransparencyNot applicable
Inputs & Outputs
- VideoInputvideo
Input video to upscale — max 495 frames (~16s at 30fps, ~20s at 24fps). A longer clip is rejected before it runs; shorten it before uploading.
- VideoOutputvideo
Upscaled output video
Parameters
- resolutionintegerdefault 960options: 720p, 1080p, 1440p, 2160p (4K)
Minimum output resolution (720p / 1080p / 1440p / 2160p (4K)) — a source already larger than this is upscaled, never shrunk
Output resolution floor (geometric mean, input aspect ratio preserved). '720p' (default 960px) is fastest; '1080p' (1440px) for full HD. '1440p' (2K, 1920px) and '2160p (4K)' (2880px) are also available — these auto-enable spatial tiling to keep GPU memory bounded (slower, more tiles; best for short clips). Higher targets need substantially more time. This is a floor, not an exact size: an input whose geometric mean already exceeds the chosen target is upscaled to its own resolution instead of being shrunk (capped at 4K), so a 1080p source never comes back at 720p.
- seedintegerdefault -1range -1–2147483647
Random seed (-1 = random)
Random seed for reproducibility. -1 (default) selects a fresh random seed every run. Set a fixed non-negative integer to reproduce a previous output.
Limitations
- May struggle with very heavily degraded footage or scenes with very fast, large movements
- On video that is already high quality (e.g. lightly compressed AI-generated content), it can occasionally over-sharpen, adding unnecessary detail
- These are described by the team as prototype models, and real-world performance may vary
Tags
Alternatives in Video Upscaling
- FlashVSR v1.1
One-step diffusion 4x video super-resolution (FlashVSR Tiny pipeline). 21 to 900 input frames (~30s at 30fps), up to 1920 px on either side.
- SeedVR2 7B
Max-quality video upscale to 720p–4K (SeedVR2 7B). Max 495 input frames (~16s at 30fps). 5-15 min (longer for 4K).
- SparkVSR
Diffusion video upscale (2x/3x/4x) with optional ref image. Max 192 input frames (~6s at 30fps). Input resolution up to 1920×1080 in either orientation.