Resized to 83% of original (view original)
| Prompt | Ignore input image. 1:1 square image, comic with three horizontal panels. Do not add extra, unprompted text or labels not quoted in this prompt. Top panel: A detailed Moogle with slight grumpy expression explains while holding a pencil: `Art time lapses faked with AI lack the classic "Draw some circles, then draw the rest of the fucking owl."` Middle panel: Label reads "Traditional". Step "1" is a small circle (represents head) on a bigger circle, and Step "2" is a finished drawing of an owl. Do not split this panel. Bottom panel: Label reads "Generated". Step "1" is a basic outline of the owl (OUTER OUTLINE ONLY), and Step "2" is the same finished drawing of an owl as traditional. Do not split this panel. |
|---|---|
| Negative prompt | Unknown |
| Model | chatgpt-image-latest (20251216), gemini-3-pro-image-preview |
| Note | Note Used a small solid grey square; "Ignore input image" is because this model requires an image input on lmarena. I manually blanked the Step 1 panels before running Prompt 2 with Nano Banana Pro. |
| Prompt 2 | The bottom left box should contain an outline sketch (outside only) very closely matching the final owl body, including the detailed feathering outline, no guidelines, no circles or simple shapes, as if the AI knew exactly what an owl looks like. The top left box should contain a drawn circle overlapping an oval, non-sketch, in the approximate size of the final owl body. |
| Prompt 3 | Reduce the bottom left drawing to OUTER outline only. The top left drawing should be single stroke circle and oval only (the input contains sketchy shapes). |

GPT: base generation asset #583843
NBP: improve step 1
Deboonkers made me think of this.
I saw a "real life" video of a physical drawing that erased a significant part, but it knew exactly what to erase, and it was already fully shaded beneath the new part. No guidelines used at any time.