sismotik21
SEEDVR2 is the absolute best AI video enhancer on the market without a close 2nd. It is free, open-source, but you need to figure out how to set it up in ComfyUI and there is a learning curve on how to tweak it to work properly with your system and videos. The output quality is unmatched. As a rule of thumb, use the 3BQ8 model, set a batch size of minimum 45 frames but preferably 61+ because the higher the less flickering and better the results. Unlike Unifab and other AI enhancers that upscales individual frames at a time, SEEDVR2 looks at multiple frames depending on batch size, compares them together for temporal consistency, allowing it to understand how the objects are supposed to look like, which produces smooth, artifact free objects and helps prevent flickering—a common issue in the typical AI enhancers like Ersgan, Unifab, Aiarty, Videopro, Vikpea, Nero, etc, which are non-diffusion-based, frame-by-frame upscaling models. SEEDVR2 also “generates” details based on its analysis of multiple frames, whereas other non-diffusion based enhancers like Unifab mainly sharpens the image. But for SEEDVR2 to work properly, you need to upscale at least 2x, preferably 3x, and 4x upscale is optimal. If you do 1x enhance, the flickering is really bad. So if you want to enhance a 720p video for example, you need to first downscale it to 360p and then do a 2x upscale back to 720p or 3x upscale to 1080p which is far better than 2x.
If you want a simple to use diffusion based AI video enhancer with temporal consistency, you can try Topaz Starlight. However, the quality is no where near as good as SEEDVR2 and runs at about half the speed of SEEDVR2.
I suggest Unifab should just integrate SEEDVR2. There will be a huge market given how most end-users wouldn’t want to spend time on how to set-up and run comfy-ui.
