I compared four models:
- SLS speed 0.6 fps
- SLM speed 0.5 fps
- SeedVR2 speed 1.02 fps
- FlashVSR speed 1.12 fps
You can view the images at this link:
https://imgsli.com/NDI1MzE3/3/0
Look at the details in the videos. SLS 4.5x seems to be the best, but it has some issues: it misidentifies inanimate objects, and human faces are still rendered artificially by SLM and SLS. It looks like a robot is in the video. SLS is best for cleaning, but SeedVR2 is best for preserving detail. FlashVSR is new and still unstable. I rewrote its code yesterday and made it more stable. It runs as a CLI on ComfyUI. Here’s the link:
A new version of SeedVR2 is coming. It’s the most promising one, and I still use it for my general videos. Here’s the link:
SLS is promising, of course, if they improve the artificial faces and text correction. Also, considering human eyes, teeth, and general features, SeedVR2 is still the best, most natural-looking version. Perhaps Topaz could maintain this natural upscaling by creating a different model in the future.
Original video:
3x SLS
4.5x SLS
4x FlashVSR Stable
4x SeedVR2
4x SLM