VibeOps

Creative Image Construction: The 9-Square Method

You remember that post I made here about JSON prompting? Well, let’s dive into a fun twist—image generation! Honestly, creating images can be super tricky because it’s tough to perfectly describe what we have in mind.

So, how about we use a method similar to taking photos? Picture this: we break the image into 9 equal squares. We’ll focus on each square one at a time and clearly describe what we want to see in each section.

For example, let’s talk about my daughter’s favorite cartoon!

ok, to get an image description let’s feed an AI (in my case I used perplexity) and then feed nano banana with my prompt

1st attempt

Sure thing! Let’s jazz it up a bit:
“Alright, alright, it seems my list may have pulled a fast one on you! Let’s toss in a keyword to spice things up!”

	"description" : "compose the image using the below list, seamless merge the elements to create a single image"

second attempt

A woman with black hair and red polka dot costume poses in front of the Eiffel Tower, holding a golden keyring.
A young woman dressed in a red polka dot costume with black accents holds a golden keyring, set against the backdrop of the Eiffel Tower and Parisian skyline.

I can’t honestly stop laughing, but well…. clothing, hair, background are matching

	  {
		"position": "Row:2 Col:2",
		"description": "Middle center square. Shows the upper part of the costume and young woman face covered by a mask; crossed hands holding a golden keyring are visible, along with the typical black polka dots on the red costume. Some object details and parts of red ribbons are seen."
	  },

see now what we generate

A character in a red costume with black polka dots stands in front of a cityscape, with the Eiffel Tower visible in the background. The character has blue hair, wears a mask, and holds a golden keyring, with red ribbons flowing around.
A vibrant character in a red polka-dotted costume with a mask poses confidently against a scenic backdrop, symbolizing creativity in image generation.

that is really way better

Note that in all the generations the model picked up the same exact character of my daughter’s cartoon, while the second generation pulled out a completely different person (even with the same elements)

one last detail

	"framing" : "close-up",

and….

A digital composite image featuring a character inspired by a popular cartoon. The character has blue hair and wears a red costume adorned with black polka dots, along with a matching mask. The image is divided into a 3x3 grid, displaying multiple iterations of the character's pose, with arms crossed and a keyring visible.
Close-up view of a cartoon character in a red and black costume, showcasing her facial features and distinctive hairstyle.

….screwed up 😦

let me rephrase

"description" : "compose the image using the below list, seamless merge the elements to create a single close up image",
A close-up image of a character with blue hair, wearing a red mask and polka-dotted costume, holding a golden keyring. The background features a city skyline.
Close-up of a character in a red polka dot costume with a mask, set against a cityscape background.

interesting, but the element positions are gone

one last attempt

"description" : "compose the image using the below list, seamless merge the elements to create a single close up image, respect the given order to place squares like pieces of a puzzle"
Image showing a grid of nine squares featuring a young woman in a red costume with black polka dots and a mask, holding a golden keyring. The background portrays a cityscape, emphasizing various expressions and angles of the character.
A close-up image of a cartoon character in a red costume with black polka dots, displaying crossed hands holding a keyring against a city backdrop.

ok I surrender

let’s try chatgpt

A cartoon character with blue hair and a red costume decorated with black polka dots, posing against a background featuring the Eiffel Tower.
Colorful close-up of a cartoon character in a red spotted costume with a blue hairstyle, set against a blurred background of the Eiffel Tower.

the best as of now

perplexity

A close-up collage image of a woman in a red polka dot costume, with a mask and blue hair, holding a golden keyring. The background shows a view of Paris, including the Eiffel Tower.
A close-up of a character in a red polka-dot costume, holding a golden keyring, set against a Paris backdrop featuring the Eiffel Tower.

perchance: random images

second attempt on chatgpt

A young girl with blue hair and a red costume featuring black polka dots, holding a golden keyring. The background shows a blurred view of a city, possibly Paris.
A close-up of a girl in a ladybug costume, featuring black dots and a red color scheme, holding a golden keyring. The Eiffel Tower is visible in the background.

foocus (yes, I am old fashioned) random images

conclusion

In theory, it sounds straightforward—describing an image piece by piece and assembling it like a puzzle. Yet, this technique often falls short with diffusion models and even older frameworks. Interestingly, it shines brighter in autoregression models!

What do you think? Dive into the conversation and share your insights!

Below the JSON I used

{ "task" : "image generation",
"description" : "compose the image using the below list, seamless merge the elements to create a single close up image, respect the given order to place squares like pieces of a puzzle",
[
{
"position": "Row:1 Col:1",
"description": "Top left square. The edge of dark blue hair is visible, a light blue sky background, and the upper part of a blurred building in the distance. No costume details present."
},
{
"position": "Row:1 Col:2",
"description": "Top center square. A significant amount of dark blue hair is visible around the top of the head, blue sky, and the beginning of red ribbons. No hands or costume details."
},
{
"position": "Row:1 Col:3",
"description": "Top right square. Background is mainly sky and a blurred Paris cityscape, with the tip of a red hair tie coming in from the side."
},
{
"position": "Row:2 Col:1",
"description": "Middle left square. There are hair, parts of the red costume with black polka dots on the left edge, and part of a red hair ribbon. The background includes a blurred city and sky."
},
{
"position": "Row:2 Col:2",
"description": "Middle center square. Shows the upper part of the costume and young woman face covered by a mask; crossed hands holding a golden keyring are visible, along with the typical black polka dots on the red costume. Some object details and parts of red ribbons are seen."
},
{
"position": "Row:2 Col:3",
"description": "Middle right square. Blurred Paris city background, sky, and a part of the shoulder with the red polka dot costume and edge of a red ribbon. No hand visible."
},
{
"position": "Row:3 Col:1",
"description": "Bottom left square. Lower part of the red polka dot costume and part of the left hand. Foreground includes the terrace edge and blurred city."
},
{
"position": "Row:3 Col:2",
"description": "Bottom center square. Shows the crossing of red hands in detail; the golden keyring held between the fingers, black polka dots on the costume, and part of the light terrace."
},
{
"position": "Row:3 Col:3",
"description": "Bottom right square. Final part of the red polka dot costume, edge of the right hand, light terrace, and blurred city background."
}
]
}

Leave a comment