Bug Report: ChatGPT cannot pass seed to Dall-E properly.

[frigid smelt]

I was able to reproduce smoething similar in my testing.

[sudden patio]

metooooo

Seed other than 5000 cannot be applied.
Also, most images published on this channel are probably seed5000.
If I copy the prompt and generate it, I can generate the same image.

The browser is irrelevant, and it occurs on both PCs and iOS applications.

#daily-theme message

#daily-theme message

If it is in the experimental stage, that's fine, but if those images are bought and sold as products, wouldn't that be a very troublesome problem?

[deft vortex]

Possibly. Actually makes ChatGPT (Dall-E) difficult to use because you simply can't get variation images from a prompt. You have to copy a prompt over to MS Image Creator to start iterating through minor variations. 😦

[fossil terrace]

Totally have had this issue too

Without even having a seed I have gotten the same images many times

[sudden patio]

No sign of a fix...is this being recognized?

[deft vortex]

🤷‍♂️

[quasi marten]

I can confirm this bug still occurs.

[grand lava]

ya I realized the same thing, minor edits would while not the best true way, would be the way to go around this

[grand lava]

the free version on bing works tho at least 🙂

[sudden patio]

It's sad that that one is not available for commercial use.

[plush furnace]

It’s a known and intended feature. As it’s in beta, it’s needed to debug and produce reliable output, not random.

[sudden patio]

Are you saying that the specifications allow you to sell the exact same image as someone else?

[plush furnace]

Huh?
I’m not sure what you intend to say.
It’s not anything to do with selling images.

What I’m saying is that the same seed is intended. That’s all.

[sudden patio]

My question is, if commercial use is allowed, would it cause a problem if the same image is sure to appear in the same situation? So, my point of view is a little different.
We should not be forced to debug without permission and for a fee.
That kind of debugging should be done in a closed environment.

As it is now, there is no system to stop people with bad intentions.

If SEED is fixed, I think the prompt absolutely needs to be confidential.

[plush furnace]

You’re using a beta product.
And I’d be very, very careful selling, possessioning, “claiming ownership of” anything “ai” generated.

Id be more worried about the real copyrights behind the sources than having a duplicate image
Also, copyright matters are different from country to country and OpenAI can’t help you resolve those. Most countries didn’t even regulate ai generated image copyrights yet.

So most likely, you’ll have much bigger issues than a duplicate image, if you sell ai generated images. Which again… I’d not do. From principle, personal pride and professional pride point of view.
But that’s your decision to make.

[deft vortex]

Can you link to where that was said? Moxi said they were actually working on fixing this and seemed to acknowledge it was an issue.

[plush furnace]

In one of the other 300 threads about this very topic.

[deft vortex]

You mean the other two bug reports?

Or do you mean the random conversations in a dozen channels?

[plush furnace]

Bug reports I believe

[deft vortex]

Hmm, I see no official response in them that this is intended or that contradicts MOxi talking about it being worked on.

[plush furnace]

I’ve read it about 2 days ago I think.
And I never read “we work on a fix”. Where is that written?

[deft vortex]

#images-discussions message

#images-discussions message

[plush furnace]

I see! Well then!
Stand corrected I guess.

[sudden patio]

Hmmm, I may be a little off topic.
I see this issue as a "can-do or can't-do" situation, not a "will-do or won't-do" situation.
But as one user, if they are working on solving the problem, all we can do is wait and see.
But it is certainly hard to recommend aggressive commercial use.
There are many forms of commercial use other than selling as is.

[plush furnace]

That is true… just was pointing out that hanging the commercial use on the identical images potentially being produced is irrelevant in terms of commercial use:
Ai generated images aren’t “your” work. Indeed all you do is input a “should look like”. It’s even less work than with a SD generator where you actually have to write a prompt.

So any commercial use of ai generated images as such, unedited, is - IMO - a weird idea anyway to start with.

I can totally see the need of a seed control. As it’s the entire “secret” to the sauce in those things, one will want to control it. But not due to commercial reasons, imo.

I strongly doubt there is even a profitable market for ai generated images. Unless of course one sells it as an awesome designers work. Which would be … well. Untrue.

[sudden patio]

I believe that AI generated (or had generated) images will have value.
Of course, now is probably not the time, but I also believe that it is not far off (not take 10 years).
Well, all that said, it's all speculation.
If it makes the world more interesting, that's OK .

Considering that the world is full of "products made by someone (or something)", it is obviously ridiculous to say that something made by a person using AI has no value.
The logic is broken.
To say that this does not convince me would be nothing more than an emotional issue.
But it won't happen while AI is still our dubious neighbor, and there are certainly mechanisms to overcome.
It will be what it will be.

[frigid smelt]

(bump for visibility)

[barren edge]

Hi y'all! I've conversed about this in #1165052598550806549 in the suggestions tab. Consider upvoting that post as well for higher likelyhood of being heard!! 💖

The problem is that they're confining users to use the same seed every single time they generate an image, meaning that they're limiting the variety of pictures that can be created. Even if asked, GPT does not allow us to change the seed. I am extremely disappointed. 😞

[frigid smelt]

Yeah, seem like API issue. GPT can explain the exact request it is issuing. Thing is that API ignores it. But it has broader implications. Similar to MJ OpenAI grants rights of generated images to users that have produced them. If the seed is same, now it is very likely that two users will have 'ownership' over same image. It is definitely not intentional and should be addressed.

[barren edge]

According to @worldly coyote they'll be changing it to a variable seed soon, but they did it purposefully for debugging. I don't feel okay with them not telling everyone about it.

[frigid smelt]

I think it is fine. These are minor bugs as these systems become actual products. OpenAI was up to few months ago a research team. GPT up to now were research projects. But they'll figure it out. Pretty awesome tech 🙂

[worldly coyote]

This bug was actually added on purpose :D #images-discussions message

[frigid smelt]

happens - I've seen much worse things in IT. There are pros and cons to that decision, but copyrightability will force seed to be random (and likely from a very large random pool - so that it is not repeatable in practical terms). Meh.. much better models are yet to come, these guys are at the frontier 🙂

[barren edge]

For about a month, until Midjourney V6 comes out and revolutionizes the entire industry.

[frigid smelt]

We'll see. Quality of image generators, same as with chat bots is measured by their ability to understand requests. MJ team has nice dataset with artistic images, but for real good quality it needs to be paired with good llm. I am not referring here to prompting image model but rather LLM embeddings that image models use to get semantically rich representation that is subsequently turned into an image. OpenAI has an edge there.

Also it is unfortunate that OA team hasn't disclosed technical inovations that came with Dall-e, so it will be harder for opensource to reproduce the model.

It is unfortunate and I dislike it, but this is the current trend. We'll see.. it will be interesting as things are very dynamic.

[barren edge]

Anyone wanna run a test with me? Let’s figure out if we’re all being given the same seed or if each user is being given a single unique seed.

To do this, paste the following prompt into GPT dall e and post a screnshot of the image you get. If the image is exactly the same, we all have the same seed. If it isn’t, then each user has a unique seed.

Use the seed 99738998. Use the following exact prompt without any adjustments: A tiny cat on a leaf

Alternatively, the seeds may be given out in groups. So, a randomly chosen group of people is confined to one seed in particular, and there are many such groups each with their own seed.

[deft vortex]

You can't specify a specific seed, it won't use it. Instead, ask ChatGPT to generate something, and then ask it what the seed used was in the returned metadata from DALL-E. It's likely 5000, pretty sure I read that's just the one they picked to assign during beta.

[plush furnace]

Seed 5000 is the only seed, it is clearly visible in the request sent to/received from Dalle by gpt-4-dalle model.
No need to overcomfirm it, it’s also been confirmed by more official sources already.