Apache 2.0 from ByteDance, weights free, nothing to accept. Self-hosting means the Hugging Face diffusers library and a single GPU, though at 7 billion parameters it wants noticeably more memory than its smaller sibling. In your browser it runs on CNAPS Studio, included in the basic plan.
Like the rest of the family it restores video in one pass rather than dozens, which is the whole engineering argument: most restoration models run the footage through the network repeatedly, and this one does not. The 7B is the larger checkpoint, positioned by its authors as the higher-fidelity option when output quality outranks turnaround.
Two costs. In CNAPS Studio expect five to fifteen minutes for a clip, with 4K sitting at the long end of that range, so this is a background job. And the quality claim over the smaller variant is qualitative rather than measured: nobody has published a number separating them. Four video upscalers sit in the CNAPS Studio catalog, and this is the one you pick when the result matters more than the wait.
What it actually is
A diffusion model trained so that a single step does the work a long denoising schedule normally does. The training was adversarial and used real degraded footage rather than synthetic damage, which is what makes one pass sufficient. Output size is handled by an attention scheme that adapts to the target resolution, which is what keeps 4K results coherent instead of tiling into visible seams.
The controls are deliberately minimal. Target resolution is a dropdown of 720p, 1080p, 1440p and 4K, defaulting to 720p, and it is simultaneously the quality setting and the cost setting. A random seed reproduces a restoration exactly. That is the entire interface.
What that leaves out is worth saying plainly: there is no restoration-strength control, no denoise amount, no way to ask for a lighter touch. You choose an output size and the model decides how much detail to rebuild, which is precisely why its own authors warn about oversharpening on clean input.
What is published, and what is not
There is no benchmark table. No restoration score, no comparison against other upscalers, and no published measurement separating this checkpoint from the 3B. The paper behind the method went up in June 2025; the model card itself carries no numbers.
The limitations section is the valuable part, and it is written by the team that trained it. The models are not robust to heavy degradation or very large motion, and may fail to remove the damage or produce unpleasant detail when pushed. On lightly degraded input they tend to over-generate detail and oversharpen. And they are described as prototype checkpoints whose real behaviour may not match the paper. Take that seriously before running a batch.
Adoption is the only outside signal, and it is modest: 9,619 downloads in the last month against 61,865 for the 3B. Six times as many people reach for the smaller one, which tells you something about where the practical sweet spot sits even without a score to compare.
How it compares to the other upscalers here
No video upscaler in the catalog publishes a comparable quality figure, so what follows is about what each one is for rather than which wins. The dividing line is restoring damage versus enlarging clean footage.
Reach for SeedVR2 7B when the source is genuinely degraded, the output is going somewhere it will be scrutinised, and minutes of processing are acceptable. Reach for SeedVR2 3B for the same job when you are working through many clips or want a faster answer, which is what most people do. Reach for FlashVSR-v1.1 when the footage is already clean and you only want it four times larger, quickly, and SparkVSR when you have a short clip and a reference frame to steer it. For a still image rather than a clip, PiSA-SR and SMFANet+ are the right tools.
What to chain it with
A video goes in and a video comes out, so this sits in the middle of a chain rather than at either end.
Two chains earn it. For generated footage, a clip from Helios Base or Helios Distilled arrives at 640 by 384, and passing it through here at 1440p or 4K is what makes it usable on a real screen. For archive work, run the restoration first and everything else after: Subject Follow (9:16) for a vertical cut, or Video Split into Image Colorization and back through Video Reassemble when the source is black and white. Restoring before reframing rather than after matters, because a crop throws away pixels this model would otherwise have used.
Open SeedVR2 7B in CNAPS Studio, run the same degraded clip here and on the 3B at 1080p, and decide whether the difference justifies the extra minutes on your own footage.
Sources
- huggingface.co/ByteDance-Seed/SeedVR2-7B for the Apache 2.0 license, the parameter count, the June 2025 publication, the limitations quoted above, the prototype status, the absence of any benchmark table, and the monthly download count.
- huggingface.co/ByteDance-Seed/SeedVR2-3B for the one-step training approach, the adaptive attention, and the smaller variant's download count.
- docs.cnaps.ai/ai-model-hub for the CNAPS Studio input, the target resolution options, the seed setting, the expected processing time, and the other upscaling models in the catalog.