New Deinterlace Model Algorithm based on Chronus

I think that a good way to deinterlace without residues or artefacts could be the follow algorithm.
Example with 720x576 sar 16:15 = 768x576 sar 1:1 → 1440x1080 sar 1:1

Load the Image
Resize with “nearest neighbor” to 720 × 288. (Important to be nearest neighbor with no aliasing filters, this will drop 1 field)
Upscale with Neural models to 768x576 (or maybe to the final resolution? 1440x1080)
Apply chronos models to double the fps and recreate the missing field in step 1.

Final result 1440x1080 sar 1:1 50fps

Please consider this model since you already have the tools to make it happen :smiley:
*Also this algorithm can be tested on the new ffmpeg version by chaining filters! :partying_face:
@suraj @gregory.maddra

The whole point of implementing VEAI as a FFmpeg filter is to allow users to chain other filters and models together. I would recommend YADIF or some other deinterlacer instead of downscaling with nearest neighbor.

bwdif=mode=1:parity=0:deint=0 is superior to yadif in my opinon.

bwdif = Motion adaptive deinterlacing based on yadif with the use of w3fdif and cubic interpolation algorithms.

The reason for nearest neighbor is to not use any upscaler that is not Veai neurons.
The real data inside the Interlaced frame are two 720 × 288 fields. the other missing 288 are being “guessed” by deinterlacers.

For example NNEDI3

I believe that taking the original 720x288 field and throwing it to veai for upscale and increase fps can give better results and less artifacts.