SeedVR
SeedVR (SeedVR2) fixes blurry videos and sharpens images in one step no fancy tuning needed
Overview
Most upscalers take dozens of steps just to clean up a video. Not this one. SeedVR does it all in one shot. No fancy loops no long waits just sharper smoother video and images straight out the gate. ByteDance teamed up with Nanyang Technological University for this.
No teacher model. No extra steps. Just smart training with adversarial learning and some custom tricks like resizing its attention based on resolution. That’s how it handles full HD without the weird glitches you sometimes get from other models.
And speed? SeedVR2 is 4× faster than SeedVR1 at 1080p.
The good stuff. It runs 15 to 50 times faster in single-step mode than older methods. Keeps time-based details consistent without extra work. Does well even with damaged or compressed footage.
The not so good. The VAE parts eat up 95 percent of the processing time. It needs a lot of VRAM even when tuned. Clean content can end up too sharp. CFG scale isn’t working right now fix is on the way.
It struggles with super compressed video or fast motion. Also if your video’s already decent it might overdo the sharpening. And that encoder/decoder (VAE) part can drag a bit on long clips.
Still it’s free it’s open-source and it’s up on GitHub right now if you wanna check it out.
Tags
Freeware Apache License 2.0 PC-based #Video & AnimationLinks
This tool is free to use when installed locally and is offered under Apache License 2.0.
Reddit folks are already stacking it up against older tools and picking it as the new go-to.
One user wrote: “For video upscaling it does require >18 GB VRAM. For images you can run it in under 10 seconds on 16 GB.”
Another tested it with a 5070 Ti and 7 B model “6.5 seconds to upscale an image of 300×300 to 1440p.” What’s cool is it’s not just for video. It’s actually become a favorite image upscaler for a lot of folks using AI tools like ComfyUI.
One Reddit user put it like this:“It’s actually also a fantastic image upscaler! It adds so much correct details to eyes skin hair fabric etc without changing the overall image.”
Another said it’s “even better than SUPIR and LDSR imo.” [ Reddit ]
A user shared an experiment using Wan 2.2 with SeedVR2 upscaling to create highly realistic AI-generated animation. They posted links to their workflow on Discord and Pastebin. The render took 15 minutes on an H100 GPU, about 30 minutes on a 5080, and potentially doable on a 4090 with longer times. They noted that the “lightx2v” LoRA is messy and that SeedVR2 adds detail without altering features, with blockswap helping reduce VRAM needs.
Feedback on SeedVR2 in that thread was mostly positive, with some technical notes:
- High detail without distortion. UAAgency said it “upscales at almost incredible level of detail” and keeps features intact.
- Blockswap helps VRAM use. Adding blockswap to the workflow lets it run on lower VRAM cards like the 5090, and maybe the 4090 if settings are tuned.
- VRAM limits still a challenge. Some users with a 4070 Super or similar found they had to drop batch size to 1 to avoid OOM errors, which reduced temporal coherence.
[ Reddit ]
Useful Links
Official SeedVR2 Video Upscaler for ComfyUI
Version
Official release of SeedVR2 for ComfyUI that enables Upscale Video/Images generation.
This page was last updated on August 13, 2025 at 1:37 AM