Post

Background: Charlie noticed this strange version of a well-known concept art piece by Ralph McQuarrie for Star Wars (1977) in the recent book about ILM (which, other than that, is apparently fantastic). So he wanted to know if it was a printing error. But it looks more like a botched upscaling job.

This post is unavailable.

I don't see why anyone would ask an AI to make a McQuarrie drawing look bad, but anything is possible in this life. Well, I guess it is? Or some weird filters? And if that's the case, it's sad that this false, slop reality is even contaminating the new official books.

This is probably a software-generated reinterpretation of the concept art. It confuses certain shapes and blends them together. In any case, I just checked my archives and couldn't find any ugly versions of this concept art.

Robert Rose@robear.bsky.social

The contrast is less of an issue than the smeary disjointed-ness of what were previously deliberate crisp straight lines. Just a different photo with different levels would not do things like break up the strut crossing over the engine section.

5 reposts44 likes

There's no room for doubt. Someone must not have had the high-resolution image, must have upscaled it, and must not have noticed that the image was very different, before dedicating a double-page spread to it. We need to change our habits when it comes to AI, and in particular: check everything.

Bill Mudron@mudron.bsky.social

Here’s the same illustration in the Kindle version. Clearly mangled by AI.

Also, you don't use Generative AI to "upscale" something. If you really need to upscale (instead of finding better sources, which there are plenty of) at least use machine learning.

That's where the whole genAI thing falls apart, doesn't it - "quick and cheap" is pointless if it requires constant second-guessing and checking by actual people. Smart companies, who care about customers and their product, will omit the "AI failing at basic tasks" bit and use real people right away

Stupid companies (ie useless management) will view the people who check and point out errors as 'difficult' and then have to go and pay other actual people a LOT more money to fix the vast problems they got from their "cheap and quick", but faulty genAI solutions from the mid-2020s.