This video demonstrates how to apply the newly added "Make Tile SEGS" in the Impact Pack to upscale using the upscale method. "Make Tile SEGS" node splits the image into tiles and creates SEGS. By using this, you can mimic the tile-based upscale function and, if the detected SEGS is too large, you can also split it for detailing. First, let's upscale the original 1024x1536 image to double its size. Simply upscale by 2x using the "Upscale Image Scale By". The result will, of course, be blurry. Let's process this image into SEGS using "Make Tile SEGS". You can see that SEGS is structured so that every part of the image can be included in the mask area. Unlike the traditional tile upscaler, this method uses Detailer, so you can improve tile heterogeneity using the 'crop_factor'. Also, setting 'mask_irregularity' to 0.7 will make the mask border irregular, improving the heterogeneity of the junctions. Next, let's separate the background and the person to alleviate the noticeable artifacts, especially in the case of humans. Apply the person ultralytics model to the "Simple Detector" to create SEGS containing the entire person. Connect the SEGS to the 'filter_out_segs_opt' in one "Make Tile SEGS" node, and in the other "Make Tile SEGS" node, connect it to the 'filter_in_segs_opt'. The node connected to 'filter_out_segs_opt' creates SEGS excluding the mask of the input SEGS, allowing you to detail the background tiles. 'min_overlap' determines how much the masks of each SEGS should overlap, and 'filter_segs_dilation' dilates the mask of the input SEGS. Increasing 'filter_segs_dilation' in 'filter_out_segs_opt' generates masks further away from the person. On the other hand, the node connected to 'filter_in_segs_opt' creates SEGS with masks overlapping the input SEGS, allowing you to detail the person. Since detailing the person requires more attention than the background, increase 'bbox_size' to avoid creating small pieces, and increase 'min_overlap' to reduce junction artifacts and allow overlapping detailing. Now, using the SEGS created in this way, let's improve the upscaled image using two Detailer nodes. Although you can handle this with "SEGS Concat", separating into two Detailer nodes allows for separate options for background and person detailing. This way, when modifying the detailing options for a person, you can prevent the recalculation of background detailing. It seems that the image has upscaled well without significant artifacts in the 2048x3072 size.
