Implementation in the STAR Project (inference_sr_img_dir_tile.py):
The script uses ImageSpliterTh from util_image.py.
ImageSpliterTh(video_data, tile_size, tile_overlap, sf=self.upscale): This line clearly shows that the tile size and overlap are configurable.
pch_size in ImageSpliterTh corresponds to tile_size.
stride in ImageSpliterTh is effectively tile_size - tile_overlap.
im_spliter.update_gaussian(output, index_infos): This method is called after processing each tile (output). Inside ImageSpliterTh, this accumulates the processed tile into a larger canvas (self.im_res) using pre-calculated Gaussian weights (self.weight).
im_spliter.gather(): This method, called after all tiles are processed, divides the accumulated results (self.im_res) by the accumulated weights (self.pixel_count) to get the final blended image. This is the weighted averaging step.
Is that possible to do tiled upscaling and be slower but not lose much quality?
can you add a script to do this for video to video super resolution upscale?
ty @CSRuiXie