How to Avoid Similar-ness?

[pliant talon]

How do I avoid it creating the same image over and over? I need variation and multiple iterations with broader differences while keeping true to the prompt. it keeps either editing the same image, or simply generating an identical one. completely opposite of what I want. happens both projects, regular and new 'images' chat.

[upper plover]

I do remember, in DALL-E3 days, we could just say "regenerate" to get variations. But with the current image generation model, we might need to approach this a little differently.

For example, I ran a quick test using this promt: "Create 3 separate images of a puppy, generated one after another. Each image includes a puppy of different breed, in a different pose, against different outdoor background."

ChatGPT generated attached 3 images, one by one, without further prompting. I don't know how it behaves with a more complex prompt or if it's able to handle more variations, but something you can try?

[pliant talon]

Doesn't seem to work for me. Neither auto, instant, nor thinking/extended make it do more than a single image. Sometimes it puts multiple images into one, like a collage...

that's all it gives me.

even your prompt doesn't give me anything more than a single image (it made a collage instead).

[upper plover]

Hmm, did you copy and paste exact wording?

[pliant talon]

yes

[upper plover]

The current image model does have the tendency toward side-by-side. The good news is, though, it listened to you re: variations.

It's tricky to know what caused these different outcomes cuz it'sall about context these days. I have the memory on, so the model could have remembered what I did in the past conversations in which I used the same "generate one after another" prompt .

Also in a separate chat immediately before, I went one by one, instead of 3 consecutive gens. My first prompt was simply "A puppy" (ofc after selecting Create Image). Second prompt: "A different puppy," abs the third, "even more different puppy."

What you can try: 1) give a thumb down to the side-by-side, 2) explain that you wanted 3 separate images, etc. (basically repeat the prompt), and 3) ask what went wrong. After that discussion, try the same prompt again within the same chat, and then, in a new chat, see if there's any difference.

[pliant talon]

Alright. I'll try that out next time I do this.

[wintry spear]

unless you're set on everything being unspecified you can give it more direction without naming the breed of dog every time. one example is giving it a region or location. also leverage constraints within the prompt. here's a quick example:
image1

[location]
Osijek, Croatia

[prompt]
puppy

[subject]
local regional dog breed (reflect typical breed appearance), natural features, 

[capture]
photorealistic

[constraints]
Avoid default depictions. Make one clear, specific choice for each: dog breed, camera distance, camera angle, lighting, setting, and moment/action. Do not center-frame by default; do not use generic studio/stock-photo composition.

[caption]
name of dog breed depicted

image2 and image3 same prompt but i used Naples, Italy instead of Osijek, Croatia for the location.

edit: i use Sora for these but got similar results fia ChatGPT Images

[undone lantern]

The problem here was it took this so literal that it gave you 3 separated images, generated one after another in sequence rather than separate renders. It did what you said by creating 3 separate images of dogs but you wanted them to be their own separate image. Is the goal to have a similar composition with different breed and background ? You could try “create a variant of this image by changing the puppy breed and the setting” make sure you do not say “variants” that will give you a picture collage vs one different render. Hope this helps - it’s just from my experience when trying to make multiverse versions of my own artwork.

[astral spindle]

What you can do also is, specially if you work with Sora and not iterative with ChatGPT is:

Prompt:

A {{breed}} dog wearing {{clothing}}

Add this lines:

breed = random.choice('Yorkie', 'Pug', 'Poodle', 'Beagle')
clothing= random.choice('Hat', 'Jacket', 'Tie', 'Glasses')

in-line:

 A {{random.choice('Yorkie', 'Pug', 'Poodle', 'Beagle')}} dog wearing {{random.choice('Hat', 'Jacket', 'Tie', 'Glasses')}}

This is the approach I take.

[astral spindle]

For a natural language in-line version, this can also be used:

Prompt:

"A dog (PICK randomly one of: Yorkie, Pug, Poodle, Beagle) wearing (PICK randomly one of: Hat, Jacket, Tie, Glasses)."

Here the trick is to use upper case letters to set focus on what needs to be done as an instruction. It's not as strong binding as the previous version.

[tough radish]

This isn’t really a randomness problem — it’s a style control problem.

Most AI tools either:
• lock into one output → no variation
• or drift too much → lose consistency

I ran into the same issue and ended up solving it by separating:
style vs subject

Instead of rewriting prompts, I define a reusable style once (colors, shapes, lighting, etc.) and then just swap the subject.

Same style → completely different outputs.

That’s how you get real variation without losing consistency.

(example attached)

[astral spindle]

Is this for GPT Image Generations?

[tough radish]

It works with GPT image generation — the toys in the example were generated with GPT.

But it’s not limited to that. The idea is model-agnostic.

The key is separating:
style vs subject

You define the style once (or extract it from an image), then reuse it while changing only the subject.

That’s how you get real variation without the model repeating itself or drifting.

I’ve been building a small system around this approach.

[astral spindle]

It is important it remains just for OpenAI models. This is not a place for everything. Be mindful about it. Rule 8 of the Server.