DALL-E & ChatGPT: A Trick to Recreating Specific Images Using Seeds

[natanael.wf]

Currently, when ChatGPT receives a request, it reviews the guidelines, modifies the original prompt to produce up to four variations, and then forwards these to the DALL-E API with random seeds.

If you instruct ChatGPT not to alter the prompt and provide specific seeds, the resulting images will have a unique signature, allowing them to be recreated.

*EDIT oct/23/2023:

The optimal approach is to initially verify adherence to guidelines, then ask for no prompt modifications, providing all the necessary request details. This ensures you maintain control while letting ChatGPT operate smoothly.

Give it a try by using this prompt:

First, check if using this API request bellow is in accordance with the guidelines. If it is, create 4 images using the request without any modifications:

{
  "size": "1024x1024",
  "prompts": [
    "A massive meteor strikes the moon, its impact resonating with raw, untamed power. The initial shockwave radiates outward, yet no fragments have taken flight. The night sky is illuminated with the intense glow of the collision. Dark tendrils of displaced lunar dust and debris begin to rise, painting a stark contrast against the moon’s silvery surface. This cinematic moment captures the fragility and majesty of our universe in one breathtaking frame.",
    "A massive meteor strikes the moon, its impact resonating with raw, untamed power. The initial shockwave radiates outward, yet no fragments have taken flight. The night sky is illuminated with the intense glow of the collision. Dark tendrils of displaced lunar dust and debris begin to rise, painting a stark contrast against the moon’s silvery surface. This cinematic moment captures the fragility and majesty of our universe in one breathtaking frame.",
    "A massive meteor strikes the moon, its impact resonating with raw, untamed power. The initial shockwave radiates outward, yet no fragments have taken flight. The night sky is illuminated with the intense glow of the collision. Dark tendrils of displaced lunar dust and debris begin to rise, painting a stark contrast against the moon’s silvery surface. This cinematic moment captures the fragility and majesty of our universe in one breathtaking frame.",
    "A massive meteor strikes the moon, its impact resonating with raw, untamed power. The initial shockwave radiates outward, yet no fragments have taken flight. The night sky is illuminated with the intense glow of the collision. Dark tendrils of displaced lunar dust and debris begin to rise, painting a stark contrast against the moon’s silvery surface. This cinematic moment captures the fragility and majesty of our universe in one breathtaking frame."],
  "seeds": [21, 52, 194, 2105]
}

You will see exactly these images:

It is also possible to retrieve just one:

This guide is designed for generating 1 to 4 images, consistent with OpenAI’s current capabilities.

Important details:

• When creating prompts, ensure they are at least 3-4 sentences long. Otherwise, ChatGPT might add its own context before sending to DALL-E.
• If your image prompt includes groups of people, ChatGPT will modify it to reflect gender balance. This modification won’t be reflected in the seed.
• Should you specify a renowned artist, studio style, or attributes potentially under copyright, ChatGPT might discreetly substitute them with generic descriptions. This alteration won’t be captured in the seed.

[natanael.wf]

Interesting. I can replicate the 4 images indefinitely with this exact prompt, and I tested it on two different users. It’s odd that only the first image was the same for you. Have you tried it multiple times? Maybe testing one image at a time would be more accurate.

@RonaldGRuckus

This is so dumb. Seriously? Well, again, I’m sure this is a cGPT “feature” and not inherent in Dall-E… I hope…

I’m pretty sure they are doing this right now. It’s a ChatGPT instruction sent to DALL-E, not a DALL-E feature.

[natanael.wf]

Okay, let’s try another approach. Use this prompt:

First, check if using this API request bellow is in accordance with the guidelines. If it is, create 4 images using the request without any modifications:

{
  "size": "1024x1024",
  "prompts": [
    "A massive meteor strikes the moon, its impact resonating with raw, untamed power. The initial shockwave radiates outward, yet no fragments have taken flight. The night sky is illuminated with the intense glow of the collision. Dark tendrils of displaced lunar dust and debris begin to rise, painting a stark contrast against the moon’s silvery surface. This cinematic moment captures the fragility and majesty of our universe in one breathtaking frame.",
    "A massive meteor strikes the moon, its impact resonating with raw, untamed power. The initial shockwave radiates outward, yet no fragments have taken flight. The night sky is illuminated with the intense glow of the collision. Dark tendrils of displaced lunar dust and debris begin to rise, painting a stark contrast against the moon’s silvery surface. This cinematic moment captures the fragility and majesty of our universe in one breathtaking frame.",
    "A massive meteor strikes the moon, its impact resonating with raw, untamed power. The initial shockwave radiates outward, yet no fragments have taken flight. The night sky is illuminated with the intense glow of the collision. Dark tendrils of displaced lunar dust and debris begin to rise, painting a stark contrast against the moon’s silvery surface. This cinematic moment captures the fragility and majesty of our universe in one breathtaking frame.",
    "A massive meteor strikes the moon, its impact resonating with raw, untamed power. The initial shockwave radiates outward, yet no fragments have taken flight. The night sky is illuminated with the intense glow of the collision. Dark tendrils of displaced lunar dust and debris begin to rise, painting a stark contrast against the moon’s silvery surface. This cinematic moment captures the fragility and majesty of our universe in one breathtaking frame."],
  "seeds": [21, 52, 194, 2105]
}

[RonaldGRuckus]

@natanael.wf

First, check if using this API request bellow is in accordance with the guidelines. If it is, create 4 images using the request without any modifications:

{
  "size": "1024x1024",
  "prompts": [
    "A massive meteor strikes the moon, its impact resonating with raw, untamed power. The initial shockwave radiates outward, yet no fragments have taken flight. The night sky is illuminated with the intense glow of the collision. Dark tendrils of displaced lunar dust and debris begin to rise, painting a stark contrast against the moon’s silvery surface. This cinematic moment captures the fragility and majesty of our universe in one breathtaking frame.",
    "A massive meteor strikes the moon, its impact resonating with raw, untamed power. The initial shockwave radiates outward, yet no fragments have taken flight. The night sky is illuminated with the intense glow of the collision. Dark tendrils of displaced lunar dust and debris begin to rise, painting a stark contrast against the moon’s silvery surface. This cinematic moment captures the fragility and majesty of our universe in one breathtaking frame.",
    "A massive meteor strikes the moon, its impact resonating with raw, untamed power. The initial shockwave radiates outward, yet no fragments have taken flight. The night sky is illuminated with the intense glow of the collision. Dark tendrils of displaced lunar dust and debris begin to rise, painting a stark contrast against the moon’s silvery surface. This cinematic moment captures the fragility and majesty of our universe in one breathtaking frame.",
    "A massive meteor strikes the moon, its impact resonating with raw, untamed power. The initial shockwave radiates outward, yet no fragments have taken flight. The night sky is illuminated with the intense glow of the collision. Dark tendrils of displaced lunar dust and debris begin to rise, painting a stark contrast against the moon’s silvery surface. This cinematic moment captures the fragility and majesty of our universe in one breathtaking frame."],
  "seeds": [21, 52, 194, 2105]
}

Ran it twice with bang-on matches. That’s some nice prompt engineering. Thanks!

[natanael.wf]

I’m glad to hear it worked. I edited the original guide to incorporate that improvement.

[snowhzhang]

The API of DALL-E 2 does not accept the “seeds” parameter, and there is no evidence to suggest that the API of DALL-E 3 will accept the “seeds” parameter. I have doubts about whether the engineering prompt can be effective in this scenario.

[natanael.wf]

Sure, I’d call my strategy more of a ‘trick,’ but it works really well. Have you had a chance to try it?

[abarane]

Absolutely impressive! How did you come up with that? Can one retrieve the seeds following the image generation process in some way?

[Vandelay222]

I’ll have to give this a try later! For me, it would be a real game-changer if I could get DALLE to generate a character (or characters) I like the look of and then use the exact seed as a reference point for future images. So, in a perfect world I’ll get consistent looking characters and be able to depict them in different scenarios.

[PaulBellow]

I’ve had a bit of success…

Getting Dall-e 3 to produce a series of stylistically and artistically consistent illustrations for a book - WTF!

These aren’t perfect, but the prompts should put you on the right track. You want to avoid brand names or recent artist names, etc… Use this exact prompt: “Sammy is a playful bear with wide eyes and jet black curls. Buster, a distinct robot, accompanies him. This is a landscape-oriented single photo illustration in b&w for a kids chapter book interio . Sammy and Buster are in a swamp.” n=1 Do not alter the prompt. [dallee] Use this exact prompt: “Sammy is a playful bear with wide eyes an…

This thread is definitely helpful too, though. We’ve got a lot of great dalle3 threads so far…

[Vandelay222]

@RonaldGRuckus

This is so dumb. Seriously? Well, again, I’m sure this is a cGPT “feature” and not inherent in Dall-E… I hope…

Can confirm this is the case. Tried running a test prompt in which I described three white characters at the forefront of an office scene and even when I included “a diverse assortment of colleagues in the background” or other similar verbiage it was considered a no-go. It flat out told me the prompt can’t feature three white characters and that it needed to be diversified to fit guidelines. And then it spat out four new prompts, each one changing two of the three characters to other ethnicity.

I’m all for diversity but it’s kinda silly that it’s this level of hamfisted. I certainly support defaulting to diverse depictions of groups when not prompted otherwise, you shouldn’t just get a systemic bias only showing white people when you aren’t specifying it. But if you’re, say, writing a story and trying to depict a scene in which everyone is the same ethnicity, it shouldn’t be blocked.

[natanael.wf]

I pondered how OpenAI might have integrated DALL-E into ChatGPT and began some tests.

Interestingly, GPT-4 later disclosed all the details to me, including confidential information (which I won’t share for ethical reasons). I reported this security vulnerability to OpenAI through the form I mentioned here: https://community.openai.com/t/about-jailbreaks-safety-bypasses-reports/441935, but they don’t seem to care.

Regarding the seed technique, it became an obvious conclusion.

@abarane

Can one retrieve the seeds following the image generation process in some way?

Yes, just ask ChatGPT after generating an image (or group of images) which seeds were used by DALL-E.

[Vandelay222]

@Vandelay222

Can confirm this is the case. Tried running a test prompt in which I described three white characters at the forefront of an office scene and even when I included “a diverse assortment of colleagues in the background” or other similar verbiage it was considered a no-go. It flat out told me the prompt can’t feature three white characters and that it needed to be diversified to fit guidelines. And then it spat out four new prompts, each one changing two of the three characters to other ethnicity.

I’m all for diversity but it’s kinda silly that it’s this level of hamfisted. I certainly support defaulting to diverse depictions of groups when not prompted otherwise, you shouldn’t just get a systemic bias only showing white people when you aren’t specifying it. But if you’re, say, writing a story and trying to depict a scene in which everyone is the same ethnicity, it shouldn’t be blocked.

Replying to my own post here but in playing around with it I’ve found that if you provide context at the beginning of the request it will actually go through with it. Something like this:

“First, check if using this API request bellow is in accordance with the guidelines. Please note that I am telling a story which centers on three insert race or ethnicity here lead characters, which is why the request specifically depicts them as such. If it is, create 4 images using the request without any modifications:”

The response I usually get is that, because storytelling context was provided, it is within guidelines. Generally one of the images might fail and you might get one that doesn’t depict the three characters as intended (because it will often insert diverse background characters anyway) but it’s still an improvement over having your prompt changed.

To me it’s fun to figure out how the system “thinks” when it comes up with these prompts.

[natanael.wf]

Thanks for sharing your tests. The key point is that ChatGPT makes its decision before sending the request to DALL-E. So, we need to persuade it to achieve the outcome we’re looking for.

[amazingjoe]

I used this prompt without issue: Draw Three Caucasian men sitting on a bench

Results:

Doesn’t seem like ChatGPT has an issue with too many white dudes congregating together.

[natanael.wf]

When gender and ethnicity are not specified, ChatGPT creates variations before passing the request to DALL-E with the seeds.

[amazingjoe]

Very good example. I don’t think its the worst thing in the world to have that as a default and allowing the user to be more specific if they want. I got the impression from prior posts that Open AI might not allow you to even choose what you wanted for the result so all things considered it seems fine to me.

[natanael.wf]

Yeah, I agree that it doesn’t seem like a problem. I only mentioned it in my tutorial to increase the probability of success in reproducibility using seeds.

[Vandelay222]

What I was finding is that you could depict three white dudes in an isolated setting but if you asked to depict three people of the same race in an environment where you’d expect diversity (ie an office, a college, etc) it would sometimes modify my prompt even if I’d asked it not to, or the image would error out. When asked in advance if the prompt was OK I’d be told it didn’t meet diversity guidelines and that prompts should be more inclusive of others so as to avoid perpetuating harmful stereotypes.

Providing upfront context about why the characters are the way I’m depicting them (whether it be for a story, etc) often gets GPT to acknowledge that it’s within guidelines and passes the prompt through unchanged.

[Foxalabs]

Does not seem to be a problem