zsxkib/seedvr2

🔥 SeedVR2: upscale ragged videos and images into sharp 4K-ready footage in a single step 🎬✨

Public
562 runs
License
GitHub
Weights
Paper
Playground API Examples README Versions

Examples

View more

Run time and cost

This model runs on Nvidia H100 GPU hardware. We don't yet have enough runs of this model to provide performance information.

Readme

SeedVR2-3B: One-Step Video & Image Restoration 🎥✨

Overview 🔊

SeedVR2-3B is a one-step diffusion transformer that restores videos and single-frame images with the exact same pipeline. ByteDance-Seed trained this beast to tackle AIGC artifacts, soft focus, mushy text, and low-light blur without patch-based hacks. I’ve wrapped their official release so you can upscale 720p YouTube rips, meme screenshots, and TikTok handhelds straight from Replicate.. no Conda gymnastics, no multi-GPU gymnastics.

Give the original researchers some love:

What you get out of the box 🧰

  • One-pass restoration: No patch stitching, no multi-step schedules—just one diffusion shot that stays temporally consistent.
  • Video and image support: Drop in an MP4/MOV or a PNG/JPG/WebP; the API figures it out.
  • Adaptive detail recovery: Fine print, faces, edge halos, and shimmering backgrounds all get dialed back into shape.
  • Sequence-parallel aware: You can bump sp_size when you move to bigger GPUs or longer clips, same as the official repo.

My quick takes & best use cases 🎯

  • Sharpening handheld 720p vlogs and wedding footage.
  • Cleaning up AI-generated “video” exports (frames that look watercolor-soft).
  • Restoring classic sports clips and meme screenshots without crunchy halos.
  • Batch fixing social media exports that got double-compressed.

I’ve stress-tested it with both video and stills from the web—scroll down for links you can preview.

Sample runs (straight from the API) 🔍

Input Output
Cat photo https://replicate.delivery/yhqm/e5UBk3qDZCzpO6edgks7jszPIqDGbLgxccTCsIfGUKi96jLrA/output_cfbd5ef8896d4794bca87f1d72f5a43e.png
Usain Bolt (Berlin ’09) https://replicate.delivery/yhqm/Dzvs9RE7QSIsPFGrjUchiQIwtVy1d3neoXPs28981BZXC5yKA/output_7813674bfa064d95a85ec33c913f811e.png
Meme screenshot https://replicate.delivery/yhqm/WedUgeslOlvmpEfxf6H3KyfNgB2TFmGRHPwaVOU8WVDCpQusC/output_5de4da98b094421a872d935b0f54e42b.png
Night city frame https://replicate.delivery/yhqm/mbVVkfGOskSWCiQkYbyHKQSnmUdn0eoky2fBZkQY3JQDLkLrA/output_cf4093493fca41a0ab38246dc54e3fdc.png
Anime-style frame https://replicate.delivery/yhqm/DJLP2CWWxHI4O98Q0UaVRW25yLfZhCNiZnrCpf02ukLsFylVA/output_73616796b6e54ef5ad9294683c61a422.png
“Ah, shit, here we go again” https://replicate.delivery/yhqm/AMGV77g4xYqbOFGaI6m4nbeC3FqCj7pO4b1TSjp0HjyCD5yKA/output_e2dc896f65fe4dadadedb3fe4ca4d788.png
01.mp4 (720p video) https://replicate.delivery/yhqm/ghPxd0cJJG66NF4R62bied6Ipyacu2qIHYMawDkL92PRfxlVA/output_d57e63381d63455a8faa10d74a887707.mp4

Tips for crisp results 🌟

  • Leave sample_steps at 1 unless you really need extra smoothing; that’s how ByteDance runs their demo.
  • cfg_scale ≈ 1.0 is a sweet spot. Higher values can oversharpen or hallucinate.
  • Keep sp_size at 1 for single A100/H100. Only bump it if you know you have sequence-parallel support.
  • Long clips? Trim to ~121 frames or shorter; the original team does the same for the 3B checkpoint.

Limitations & caveats ⚠️

  • Heavy motion blur or super-dark footage can still break it—SeedVR2 is good, not magical.
  • The 3B checkpoint fits on 80 GB GPUs, but 4K output still demands proper VRAM headroom.
  • It can overcook very clean footage (makes it look too sharp); dial down CFG if you notice that.

Terms of use 📚

The SeedVR/SeedVR2 projects are Apache 2.0. ByteDance notes these models can over-generate detail and aren’t bulletproof against extreme artifacts. Don’t ship anything malicious or misleading; keep it above board.

TL;DR ethics: Don’t impersonate, defame, or use the outputs for scams or political persuasion. You’re responsible for how you deploy this.

Credits 🙏

  • Core research & weights: Jianyi Wang, Shanchuan Lin, Zhijie Lin, Yuxi Ren, Meng Wei, Zongsheng Yue, Shangchen Zhou, Hao Chen, Yang Zhao, Ceyuan Yang, Xuefeng Xiao, Chen Change Loy, Lu Jiang.
  • Original GitHub: ByteDance-Seed/SeedVR

I just made it behave nicely on Replicate.

⭐ Star the repo on GitHub!
🐦 Follow me on X/Twitter: @zsakib_
💻 More projects: github.com/zsxkib