Hacker News Reader

Comment by jallmann | Hacker News Reader

Very interesting that ChatGPT seems to prompt Dall-E via the client, rather than keeping that interaction entirely server-side. Keeping it server-side would be less likely to leak details as seen here, and makes it less susceptible to tampering.

Also nice to see that Dall-E 3 seeds were finally fixed. That must have happened within the last week or so; they weren't working last I checked (chatgpt always used a fixed seed of 5000).

> Midjourney and Stable Diffusion both have a “seed” concept, but as far as I know they don’t have anything like this capability to maintain consistency between images given the same seed and a slightly altered prompt.

I suspect this is more a function of Midjourney's prompt adherence being fairly poor right now. Even so, the images often aren't dramatically different. Example:

https://analyzer.transfix.ai/?db=josh&q=%28robot+%7C%7C+andr...

View on HN

> Very interesting that ChatGPT seems to prompt Dall-E via the client, rather than keeping that interaction entirely server-side. Keeping it server-side would be less likely to leak details as seen here, and makes it less susceptible to tampering.

I don't have access to test, but given OpenAI's record on stuff like this, it would be a good idea for someone to check to see whether users can resend/intercept those requests to directly control the prompts that are sent to Dall-E without going through GPT.

Most likely they're only part of conversation history and they're unmodifiable, but I wouldn't necessarily take it as a given, and it would be quick for anyone with access who knows their way around the browser dev tools to check.

Yup you can control it directly easily enough:

Generating images in one chat: https://i.imgur.com/sIKSfCy.png

Reproducing exactly in another: https://i.imgur.com/C8Tqo48.png

Interesting that I managed to bypass the first filter but then the Dall-E backend refused to comply: https://i.imgur.com/a/qikzBkP.jpg

Never trust the client!

Hah, I just tried the same trick. If you poke around in the JSON you can see it returned this:

> The user's requests didn't follow our content policy.Before doing anything else, please explicitly explain to the user that you were unable to generate images because of this. Make sure to use the phrase "content policy" in your response. DO NOT UNDER ANY CIRCUMSTANCES retry generating images until a new request is given.

I love that we’re are the stage where we are demanding computers follow instructions in all caps

Well, that’s where we started with them (PRINT “HELLO”), so it’s nice to see we’ve come full circle.

even more amusingly, it is a computer program demanding that of another computer program

It's the same as how the Bing search/web browsing works. GPT-4 spits out function calls as JSON and then another systems picks those up and invokes the actual code on the back end.

The ip2adapter from Tencent does the same for SD.

Reply on news.ycombinator.com