Quality Help

[Floribela]

I try using all the different models into this vídeo, and i didin't get a good result... how can i get the best result here? im stuck into this video

WhatsApp.Video.2025-12-13.at.19.38.39.mp4

Original

3B

other seed

other seed and other resolution 4x

7b with other resolution

7b sharp

3bq8

[urixl]

First of all, downscale your video.
As we all can see, there's not much details in the original video due to compression, so we won't lose any details at all.
I'm downscaling good SD videos for SeeedVR to work on.
I've downscaled your video to 382x288 and upscaled it at 3x ratio with the 3B Q8 model.

See the result:
https://drive.google.com/file/d/1kbxEFFxIdEitdz9t4Jk80Ik0j61rdFVh/view?usp=sharing

And one more thing: don't judge your result by the first frame. It'll always be the worst because of the model design.
The more frames it'll analyze, the better-looking result will be. So I'm using batch_size of 73 frames and the most refined frame is the last.
Unfortunately, this modes doesn't use look-ahead algorithm to predict what frame will look.
It uses statistic from previous frames only. Yes, it can stitch batches with 8 or more frames to eliminate dramatic jumps in frame quality.
But the first frame of the video will always be the worst.
You can set the prepend_frames option to the same 8 frames, for example, to avoid such quality loss.

[urixl]

And to add some finishing touch: Topaz Starlight after SeedVR, then Topaz Apollo - 30fps, added some noise to mask obvious blank areas.

Download the result:
https://drive.google.com/file/d/1UapzbxAIWCSayCx8ImCttZnGELa8KPQ0/view?usp=sharing

Now, let's compare with the original: