Faking a JPEG

#1

Post by todsacerdoti » Fri, Jul 11, 2025, 10:57 PM UTC

Re: Faking a JPEG

#2

Post by lblume » Fri, Jul 11, 2025, 11:43 PM UTC

Given that current LLMs do not consistently output total garbage, and can be used as judges in a fairly efficient way, I highly doubt this could even in theory have any impact on the capabilities of future models. Once (a) models are capable enough to distinguish between semi-plausible garbage and possibly relevant text and (b) companies are aware of the problem, I do not think data poisoning will be an issue at all.

Re: Faking a JPEG

#3

Post by bschwindHN » Fri, Jul 11, 2025, 11:51 PM UTC

You should generate fake but believable EXIF data to go along with your JPEGs too.

Re: Faking a JPEG

#4

Post by mrbluecoat » Fri, Jul 11, 2025, 11:52 PM UTC

> I felt sorry for its thankless quest and started thinking about how I could please it.

A refreshing (and amusing) attitude versus getting angry and venting on forums about aggressive crawlers.

Re: Faking a JPEG

#5

Post by ASalazarMX » Fri, Jul 11, 2025, 11:56 PM UTC

mrbluecoat wrote: post #4

> I felt sorry for its thankless quest and started thinking about how I could please it. A refreshing (and amusing) attitude versus getting angry and venting on forums about aggressive crawlers.

Helped without doubt by the capacity to inflict pain and garbage unto those nasty crawlers.

Re: Faking a JPEG

#6

Post by dheera » Sat, Jul 12, 2025, 12:08 AM UTC

> So the compressed data in a JPEG will look random, right?

I don't think JPEG data is compressed enough to be indistinguishable from random.

SD VAE with some bits lopped off gets you better compression than JPEG and yet the latents don't "look" random at all.

So you might think Huffman encoded JPEG coefficients "look" random when visualized as an image but that's only because they're not intended to be visualized that way.

Re: Faking a JPEG

#7

Post by EspadaV9 » Sat, Jul 12, 2025, 12:11 AM UTC

I like this one

https://www.ty-penguin.org.uk/~auj/spigot/pics/2025/03/25/fa...

Some kind of statement piece

Re: Faking a JPEG

#8

Post by derektank » Sat, Jul 12, 2025, 12:22 AM UTC

bschwindHN wrote: post #3

You should generate fake but believable EXIF data to go along with your JPEGs too.

From the headline that's actually what I was expecting the link to discuss

Re: Faking a JPEG

#9

Post by jesprenj » Sat, Jul 12, 2025, 12:28 AM UTC

lblume wrote: post #2

Given that current LLMs do not consistently output total garbage, and can be used as judges in a fairly efficient way, I highly doubt this could even in theory have any impact on the capabilities of future models. Once (a) models are capable enough to distinguish between semi-plausible garbage and possibly relevant text and (b) companies are aware of the problem, I do not think data poisoning will be an issue at all.

Yes, but you still waste their processing power.

Re: Faking a JPEG

#10

Post by hashishen » Sat, Jul 12, 2025, 1:03 AM UTC

the hero we needed and deserved