How do I make DALL-E 3 generate a reliable grid (3x3) of images?

Friendly greetings! In ChatGPT, I’m trying to make DALL-E generate some flipbooks, or animation-esque sequences.

Sometimes, it works. But often, the grid is off-kilter, or it’s a different number of squares then what I specify, like they’ll be rectangles, or it’ll be 4x3 etc. I’ve asked for things like “9 squares”, “9 cells”, and so on and so forth, but it will quite frequently disobey or deviate from this literal and logical prompt.

Is there a better way to word this that’s more consistent? Would appreciate any pointers.

I’m trying prompts that are both very to-the-point and a lot more verbose. Here’s an example of where it clearly gave me a 4x3 with unequal cells, AND white space.

But if I re-run it in a new session… this is what it gives me, which is more of what I intend. I just can’t seem to achieve consistency.

Yeah, in my experience it’s not 100% there yet for stability. Maybe DALLE4? It’ll get better soon.

1 Like

You’ve been frequently trying grids too?

Does anyone else have similar experiences?

There’s really no “reliable”. The AI seed plus your prompt either gets oddball grid, or not, on successive runs. Even sometimes drawing lines through or partially through images.

Resubmits with the same prompt and gen_id:

Sending again immediately after without the gen_id:

(which now doesn’t follow the “diverse individuals” prompt as well)

Using a method to ensure the prompt is unaltered.

The progress of a story through the panes is also very hard to get followed.

I think the wide format vs square gives the AI more challenge, as evidenced from other symptoms like the white side borders above, it seems to use some outfill technique to reach wide.

More random grid fun…that was not specifically instructed, nor was the content directly instructed.