SparkVSR — Video Upscaling AI Model
SparkVSR upscales video 2×, 3× or 4× and will take a high-resolution still as a reference for a chosen frame, so an image model can set the look the whole clip follows. Up to 192 frames in — about 6 seconds at 30fps, 8 at 24fps — so shorten a longer clip before you upload it. Accepts video up to 1920×1080 in either orientation, so 1080×1920 portrait clips pass too. With a reference image the whole clip is processed in one pass, which lowers the frame limit for large outputs (about 57 frames for 1080p at ×4); a longer clip is rejected with AI.OVER_MAX_FRAMES that reports the exact limit.
What it does
SparkVSR upscales video 2x, 3x or 4x, and unlike most upscalers it lets you steer the result. Give it one already-upscaled still and the index of the frame it came from, and it carries that look across the clip while following the motion of the original footage.
That is the way out of an odd-looking face or texture: fix a single frame with whichever image upscaler is sharpest, then let the video model follow it. Leave the reference empty and it runs blind, as an ordinary upscaler. Built on CogVideoX 1.5 by researchers at Texas A&M and Google.
When to use it
- Upscaling 2x, 3x or 4x with a say in how the result looks
- Fixing an artefact by upscaling one frame yourself and passing it in
- Borrowing a sharper image upscaler's output as the reference look
- Blind upscaling, for when there is no reference frame to hand
- Image-model sharpness in a video, without frame-by-frame work
Details
- CategoryVideo Upscaling
- Year2026
- LicenseApache-2.0
Compliance & Provenance
- ProviderOpen-source · Specialized
- EU AI Act RiskMinimal Risk
- Art. 50 TransparencyNot applicable
Inputs & Outputs
- VideoInputvideo
Input video to upscale — max 192 frames (~6s at 30fps, ~8s at 24fps). A longer clip is rejected before it runs; shorten it before uploading. Accepts video up to 1920×1080 in either orientation, so 1080×1920 portrait clips pass too. With a reference image the whole clip is processed in one pass, which lowers the frame limit for large outputs (about 57 frames for 1080p at ×4); a longer clip is rejected with AI.OVER_MAX_FRAMES that reports the exact limit.
- Reference ImageInputimageoptional
Optional HR reference for the first frame (e.g. PiSA-SR or Nano-Banana output). Improves perceptual quality.
- VideoOutputvideo
Upscaled output video
Parameters
- upscaleintegerdefault 4options: 2, 3, 4
Upscale factor (2/3/4)
Enlargement factor on each dimension. 4 (default) for maximum upscaling. 3 intermediate. 2 fastest. Higher factors significantly increase compute and VRAM.
- ref_frame_indexintegerdefault 0range 0–191advanced
Reference-image frame index (0-191)
Zero-based index of the input video frame that the optional reference image corresponds to (commonly 0 = first frame). Ignored when no reference image is wired in. Use a non-zero index when you've pre-generated a sharp version of a specific later frame.
- seedintegerdefault -1range -1–2147483647advanced
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
- Input is capped at 192 frames at up to 1920x1080 in either orientation - about 6 seconds at 30fps, 8 at 24fps
- One reference still per run, applied at the frame index you choose
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 3B
Video upscale to 720p–4K (SeedVR2 3B). Max 495 input frames (~16s at 30fps). 2-5 min (longer for 4K).
- SeedVR2 7B
Max-quality video upscale to 720p–4K (SeedVR2 7B). Max 495 input frames (~16s at 30fps). 5-15 min (longer for 4K).