Image printing - the resolution question

David Pogue did some A-B testing of 13MP images resized to 5MP and 8MP, and nobody can tell the difference:

Pogue’s Posts - Technology - New York Times Blog

This is not a big surprise to me, because we've done tests like these over the past 5 years with Picasa's print engine.

Picasa does Lanczos-3 interpolation to 1/2 the native printer resultion by default when it prints, and a user can request that we use Lanczos-8 at full-res! Needless to say, I think our output looks rather good, and we do better with small-sized images than almost any other app out there.

Many major online print providers still do awful interpolation, and so part of this equation is: what kind of interpolation does the printer driver (or RIP) use when it's printing? I've seen many well-known services do nearest-neighbor resampling, and almost all consumer printers do this. You might not be able to tell, but you probably could if you looked closely.

When we chose a default "low-res" size for Picasaweb, we actually picked 1600x1200 based on some double-blind prints at 8x10.

I'm not kidding: people could not tell the difference between a 2Mpix image printed with great interpolation at 8x10, compared with an 8Mpix image printed at full resolution.

Sounds crazy, but there it is...

2 comments:

  1. Anonymous1:43 PM

    Ahhh!!! Lanczos-3!!! are you CRAZY? what about those negative lobes??!?

    On the other (er, serious) side - yeah, people generally like "soft". Which has always made me wonder... when looking at image compression and perception - there's a reconstruction/sampling vs. compression ratio graph that I've always wondered about...

    that is, are you better off with more compression or more resolution - meaning, would you get a smaller (file size) picture of some X quality by (a) compressing less at lower res and upsampling nicely on decompress, (b) compress "at res" at some medium setting or (c) compressing a LOT at higher res and downsampling nicely at decompress?

    JPEG's got a lot of artifacting....

    ReplyDelete
  2. Most bicubic resamplers have negative lobes - no problem at all having a better kernel!

    A Lanczos filter actually does a good job at edges and upsampling - mad good sharpness and beautiful smooth images when you upsample.

    Without negative lobes, you just aren't going to get anything but BLUR. For example, even the widest b-spline will just make edges blurry, much blurrier than they need to be.

    My resampler, of course, is programmable with about a dozen different kernels.

    ReplyDelete