{"id":54069,"date":"2023-10-02T22:47:10","date_gmt":"2023-10-02T20:47:10","guid":{"rendered":"https:\/\/www.burks.de\/burksblog\/?p=54069"},"modified":"2023-10-02T22:47:11","modified_gmt":"2023-10-02T20:47:11","slug":"shorten-und-nicht-die-tasten-belauschen-lassen","status":"publish","type":"post","link":"https:\/\/www.burks.de\/burksblog\/2023\/10\/02\/shorten-und-nicht-die-tasten-belauschen-lassen","title":{"rendered":"\/Shorten und nicht die Tasten belauschen lassen"},"content":{"rendered":"

\"AI<\/a><\/p>\n

Bei Fefe<\/a> las ich \u00fcber A Practical Deep Learning-Based Acoustic Side Channel Attack<\/a> on Keyboards<\/i>:<\/p>\n

When trained on keystrokes recorded by a nearby phone, the classifier achieved an accuracy of 95% […] When trained on keystrokes recorded using the video-conferencing software Zoom, an accuracy of 93% was achieved.<\/i><\/p>\n

Das ist eine ziemlich abgefahrene Sache. Mir fiel dazu Markus Kuhn<\/a> ein (der vor mehr als 20 jahren der Log Angeles Times nette Dinge \u00fcber mein Tron-Buch<\/a> gesagt hat). Kuhn hat vor zwei Jahrzehnten schon Angriffe auf devices<\/a> geschildert, die sich u.a. die Strahlung des Monitors zunutze machen.<\/p>\n

\"AI<\/a><\/p>\n

Nat\u00fcrlich wollte ich das Thema angemessen bebildern, aber das war gar nicht so einfach. Neu ist bei Midjourney der sehr n\u00fctzliche Befehl \/shorten<\/i>. Damit l\u00e4sst man die KI den Prompt, den man zusammengestoppelt hat, in token<\/i> zerlegen. Der Rest kann in die Tonne.<\/p>\n

Mein Ausgangsbefehl war risks and dangers of the dangerous video conference software zoom, symbolic photo, –stylize 500 –ar 16:9<\/i>. Den hat die KI auf Versionen von dangerous video conference, photo<\/i> gek\u00fcrzt – der Rest w\u00e4re ohnehin gar nicht ber\u00fccksichtigt worden.<\/p>\n

\"\/shorten\"<\/p>\n

What’s new in V5.2?<\/p>\n

New Aesthetic System<\/b>
\nImproved aesthetics and sharper Images<\/b>
\nSlightly improved coherence and text understanding<\/b>
\nIncreased diversity<\/b> (sometimes you may need to roll more than once to get what you want sorry!)<\/p>\n

The –stylize command<\/b> has been fixed to have a strong effect on the amount of the stylization applied to your image (more like in V3). This command goes from –stylize 0 to –stylize 1000 and the default value is –stylize 100<\/p>\n

New „High Variation Mode“<\/b>
\nTurned on by default and makes all variation jobs much more varied.
\nTo toggle this type \/settings and click a different variation mode.
\nUnderneath all upscales you can also chose which strength of variation you want.<\/p>\n

New \/shorten command<\/b>
\nThis lets you ‚analyze‘ a prompt and get suggestions on what words might not be doing anything and which ones might be key<\/p>\n

New „Zoom Out“ Feature<\/b>
\nAll upscales now have ‚zoom out‘ buttons underneath you can use to reframe the image
\n[Zoom Out 1.5x] [Zoom Out 2x] pulls the camera out and fills in details on all sides
\n[Make Square] adds details on two sides to make a non-square image into a square one
\n[Custom Zoom] (advanced) gives you a popup text box that allows you to change the prompt as you zoom out as well as the aspect ratio or precise zoom. To change aspect ratio set –zoom 1 then change the –ar to your target. This option will try to remove black bars on your previous images.<\/p>\n

\"AI<\/a><\/p>\n

Bei derart vagen Vorgaben darf man sich nicht wundern, dass das Ergebnis wie eine Mischung aus einer Zombie-Apokalypyse und einem Treffen weiblicher Taliban aussieht. Aber lustig ist es schon.<\/p>\n

\"AI<\/a><\/p>\n

\"image_pdf\"<\/a>\"image_print\"<\/a><\/div>
<\/div>