At first I was also attracted to Refile’s on-the-fly processing. However, there are multiple downsides with this:

• It would be great if the client can choose the size, however *you mustn’t expose the routes to the client, otherwise people can easily DoS you by requesting all kinds of image dimensions. That means that only the server can generate URLs, which lowers the flexibility for apps with separate backend and frontend.
• The “first” image download gives bad user experience, because the processing can take a long time (e.g. image has to first be downloaded from S3, processed, and then sent again over the wire).
• What is the right way to cache the images? On Refile people were complaining that Amazon CloudFront caches the images only for couple of hours. This is really a problem, considering the previous bullet. Caching with Nginx is also not ideal, because then you have to have large disk space on your server.

However, Refile’s uploading on file select (through the mounted Sinatra app) is awesome, that’s the first time someone made something like this. But I feel the upfront processing is still the right way, just background processing should be more convenient.