How to get gemini to follow my instructions to create this comic image?

[NewDae]

Using the Gemini model in AI studio is not great both in the output (sticking to desired aspect ratio, consistency of image, following instructions). I do get marginally better output if I keep asking it to correct itself, but then it seems like an endless loop to get the final output.

My doubt is whether I am trying to do something too advanced with the model or if something is off in my prompt. Looking forward to more experienced folks helping here!

Attached my prompt below for reference,

I absolutely need you to focus and ensure all details below are followed to make the following comic :

- [Format] 3 panels, stacked vertically (4:5 aspect ratio, 1080x1350)

• [Style] Bold black outlines, flat retro colors, dark-humor webcomic
• [Setting] Seaside café table, ocean horizon in background. Background must stay identical in all panels
• [Characters]
• [Percy] Penguin, round, nervous/anxious, expressive, often sweating. Always on the LEFT
• [Clawson] Crab, small, smug, confident, sarcastic, sipping coffee. Always on the RIGHT

- [Text Rules]

• Bold comic font
• Speech bubbles slightly larger than normal with padding + readable line-height
• Dialogue must be used EXACTLY as written (no typos, no rewording, no shortening, no omissions, no swaps)
• If text does not fit, adjust bubble size, never alter the text
• Preserve dialogue ORDER per panel exactly as provided

-–

Panel 1

- [Shot] Wide

- [Scene] Percy relaxed, sipping soda with a straw. Calm face. Clawson leans smugly with coffee mug. Ocean + sky behind them

- [Emotions] Percy = casual. Clawson = teasing

- [Dialogue Order Rule] First bubble = Clawson, no Percy bubble in this panel

- [Dialogue]

• [Clawson]: “Hey Percy, what’s common between Equifax and Facebook?”

-–

Panel 2

- [Shot] Medium

- [Scene] Percy smirking, straw in beak, sarcastic. Clawson smirks knowingly, one claw raised

- [Emotions] Percy = cocky/snarky. Clawson = smug/authoritative

- [Dialogue Order Rule] First bubble = Percy, second bubble = Clawson

- [Dialogue]

• [Percy]: “Easy. Facebook’s where my aunt spams minion memes… and Equifax? Fax machines, right?”

• [Clawson]: “Nope. Equifax is a credit bureau. And both had massive data breaches.”

-–

Panel 3

- [Shot] Close-up

- [Scene] Percy freaking out, sweating, eyes wide, clutching soda cup. Clawson in background, relaxed, sipping coffee

- [Emotions] Percy = panicked. Clawson = calm

- [Dialogue Order Rule] First bubble = Percy, second bubble = Clawson

- [Dialogue]

• [Percy]: “WAIT. My data was leaked?!”

• [Clawson]: “You can check with Mozilla Monitor — it tells you if your info leaked.”

-–

[Pannaga_J]

Hi @NewDae
Your prompt is incredibly detailed and well-thought-out; however, it requests the model to perform several cutting-edge tasks simultaneously, which typically requires a more structured, multi-step workflow. I recommend you to create three separate, highly-focused prompts one for each panel first. This allows the model to concentrate solely on generating a consistent scene, style, and character look for each single image.
Just to confirm you are using nano banana model right ?

[NewDae]

Hey @Pannaga_J

Thanks for the reply. Took a break and figured this out - one panel at a time is definitely the way to go. It worked with Nano Banana.

However, the only problem I run into is typos and then difficulty in correcting typos. Do you have any tips on how to prompt better to correct spelling mistakes?

[Pannaga_J]

I think to get 100% accuracy, you could generate the final image with empty speech bubbles and then add the dialogue in post-processing part.

If you want try to generate the text directly with the model, please try these workarounds to improve chances:
Isolate the text by enclosing the dialogue in quotation marks at the start of your prompt. You could also employ negative prompting strategies and specify a quality constraint, using phrases such as: “The text must have photorealistic typographical accuracy.”