We first introduced auto-detect masking functionality in Sharpen AI v3.3. When enabled, users have the option for Sharpen AI to auto-detect—and mask—general subjects, landscapes/skylines, and portraits. With Sharpen AI v4.1, we’ve made significant improvements to the overall detection quality of these models, as well as improved the edge detection of the auto-drawn masks.
All three auto selection models have been updated with a recent state-of-the-art deep learning architecture. In more technical terms, segmenting out the salient object(s) in an image requires us to classify each pixel as salient or not salient. Now, when classifying each pixel, this new deep learning architecture sees larger local regions to improve this classification accuracy.
In other words, this update enables more accurate masks, particularly on the boundaries of objects. Users can now better rely on auto selection, and may need less manual modifications to get a perfect mask. Here are some comparison images that showcase the improvements users can expect when moving from the original auto-detect mask models to the improved ones.
Known issues with the improved auto-select mask models
When using these updated models for the first time, some Apple computers with M1 silicon may experience longer processing times (expect 20-30s) to load the models. However, after the first use, all subsequent uses of these auto-detect models will be much faster. We hope to release a patch to reduce this initial time cost in a future Sharpen AI update.
This update is also not compatible with CPUs that do not include AVX support. For Apple computers, this can include machines manufactured prior to 2012; For Windows computers, you will need to check with your CPU manufacturer to determine AVX support. For computer that do not support AVX, Sharpen AI will utilize the previous mask models.
When new users launch Sharpen AI, they will be greeted with an option to batch download all models, making it easier and more convenient to use the application. Once a batch download begins, an unobtrusive progress bar will display the progress of the download. You can also force a batch download by clicking on Help / Install Missing Model Files.
We also made it easy to determine which models still need to be downloaded by displaying the model names in italics, as illustrated below. You’ll know that all models have been downloaded and are available locally if none of them are listed in italics.
In addition to the improvements listed above, we also make it a point to release each new update with bug fixes that make usability of our apps more reliable and better. One of the changes we made has to do with the way you save JPEG files. Instead of using a Quality scale of 1 - 10, we now allow you to set the discrete 1:1 quality amount on a scale of 1 - 100.
We’ve also improved the the brush controls while in masking mode. Now, when you right-click, the opposite brush action will be made. So, if you’re in “Add” mode and you right-click while making a stroke, it will be done in “Subtract” mode.
Finally, in addition to a number of under-the-hood bug fixes, we also improved Photoshop smart object support for auto-select masking options (excluding Custom). If you convert a layer to a Smart Object and send it to Sharpen AI, we will preserve the auto selected masking option you choose, whether it’s Subjects, People/Portraits, or Skylines/Landscapes. NOTE: custom masks are not currently preserved via Smart Objects.