Hacker News Reader

Comment by althea_tx | Hacker News Reader

Are you sure the seed number isn’t a hallucination? I think it might be an internal reference which vaguely alludes to the original prompt.

The seed has no persistent meaning beyond the chat instance, so I think you could get same effect by referring a previous image with prose.

View on HN

The seed isn't a hallucination, I proved that to myself by reviewing the underlying JSON, see this section: https://simonwillison.net/2023/Oct/26/add-a-walrus/#peeking-...

    {
      "prompts": [
         "Photo of two Muppet characters: a pelican with a monocle and a bow tie, and a walrus with big, goofy tusks and a dapper bow tie. They're seated in a Muppet-style commentary booth, providing humorous commentary on the Monaco Grand Prix. Cartoonish F1 cars race by, and colorful yachts are seen in the distance."
      ],
      "size": "1024x1024",
      "seeds": [1379049893]
    }

It's definitely the true params! https://ibb.co/jJPm7Bq

I thought the same, it seems strange it would be added to that json as a key separate to the prompt though

Why? That's just how the image generators technically work.

Because it seems to me like poor engineering to leak it. Not disputing it, just surprised to see it

Maybe I’m missing something. What is the threat of exposing the seed? It just makes images reproducible.

Remember that just like with every other model they expose, DALL-E 3 API is stateless. GPT-4 to DALL-E 3 hand-off is implemented client side (similar to plugins), so the entire exchange goes through your machine. Everything - seed included - starts "leaked" by design.

Reply on news.ycombinator.com