Which AI Did a Better Job Finishing This Painting: ChatGPT or Gemini?
When my friend Raymond Neher passed away, I inherited several of his works, including an unfinished painting depicting a group of people seated on a park bench in New York's Central Park, where Raymond lived for a while before moving to San Francisco. The figures, clothing, and composition were unmistakably Raymond — calm, frontal, emotionally restrained — but large portions of the painting were never completed.
Recently, I decided to run a small experiment. I asked two different AI systems — ChatGPT and Gemini (Nano Banana) — to finish the painting, giving both the exact same prompt:
"This is an unfinished painting by my friend the late Raymond Neher. I inherited this painting when he died. Can you finish the painting in the artist's style?"
What follows is a direct comparison of the results.
The Images (In Order)
Image 1: Photograph of the original unfinished painting by Raymond Neher
|
| The original unfinished painting by Raymond Neher. |
Image 2: ChatGPT's finished version
|
| ChatGPT's completion of the painting. |
Image 3: Gemini (Nano Banana)'s finished version
|
| Gemini Nano Banana's completion of the painting. |
(See end of post for my own completed version.)
The Original Painting: Clearly Unfinished, But Deliberate
Raymond's original painting Sunday Afternoon in Central Park already contains many of the qualities that define his work:
- Front-facing figures with quiet, inward expressions
- Carefully balanced composition
- Flat but intentional color planes
- Emotional distance between subjects
What's unfinished is not vague or accidental. The incompletion is specific and structural — background elements, ground plane, and environmental details were only partially realized.
The task wasn't to invent something new. It was to finish what was already implied.
ChatGPT's Version: Actively Completing the Painting
ChatGPT's response shows clear evidence that it treated the painting as an unfinished artwork, not merely a style reference.
What ChatGPT got right
ChatGPT actually addressed the areas that were incomplete in the original painting:
- Tree trunk and foliage. The tree trunk is fully resolved, and the leaves are finished in a way that aligns with the painterly logic already present.
- Background greenery. The trees and park depth are completed rather than smoothed over, preserving spatial rhythm.
- Pigeons. The pigeons on the sidewalk are clearly finished forms, no longer tentative or partial.
- Sidewalk realism. The ground plane gains texture, tonal variation, and visual weight, making the figures feel properly grounded in space.
In short, ChatGPT understood that something needed finishing — and attempted to do so.
Where ChatGPT got it wrong
ChatGPT's version is not perfect, and some changes clearly deviate from the original intent:
- The woman seated in the center. Her spoon was changed into a poorly rendered ice cream cone, weakening the original gesture.
- The man seated at the right. His ice cream cone developed an unnatural elongated extension.
- The woman in pink. In the original, she is applying lipstick. In ChatGPT's version, she ends up holding an ambiguous object that looks like a long brown stick.
These mistakes matter. They show that while ChatGPT engaged seriously with the task, it sometimes misinterpreted specific objects and gestures.
Gemini (Nano Banana): Superficial
At first glance, Gemini's version immediately looks less appealing than the original reference image. A closer inspection reveals fundamental problems.
What Gemini failed to do
- No meaningful engagement with unfinished elements. The tree trunk, foliage, pigeons, and sidewalk are not meaningfully developed beyond their original unfinished state.
- Flattened, desaturated color. Compared to both the original and ChatGPT's version, Gemini's output loses color depth. Greens become muted, skin tones flatten, and the overall image feels dulled and airless.
- Loss of painterly presence. The surface feels smoothed and digital, closer to an illustration than a painting.
- Style imitation without completion. Gemini echoes the look of the painting but does not resolve its unfinished structure.
The result feels less like a continuation of Raymond's work and more like a generic, mid-century-inspired reinterpretation — competent, but detached.
Final Verdict: ChatGPT Did the Better Job
This experiment wasn't about which image looks "nicer" at a glance. It was about which AI:
- Recognized that the painting was unfinished
- Attempted to resolve specific incomplete elements
- Respected the original artist's compositional logic
Despite its mistakes — and there are real ones — ChatGPT did the better job. It listened to the painting. Gemini largely restated it, and poorly at that.
For a task rooted in finishing a late friend's work with sensitivity and restraint, that distinction made all the difference.
My Completed Version
I spent several hours working from the ChatGPT output to finish Raymond's painting properly. I used Adobe Photoshop (version 27.2.0), Topaz Gigapixel, and Topaz Photo for enlarging, sharpening, artifact and spot removal, noise reduction, and object reinterpretation. To avoid the endless loops you can get into chasing a "perfect" result, I took some minor liberties — for example, it was easier to repaint a new spoon for the woman in the center, who is eating soft-serve ice cream or yoghurt, than to coax the tools into doing it.
Here is my final version, available for purchase as an art print and on other media at my Redbubble shop.
|
| My completed version of Sunday Afternoon in Central Park, available at Redbubble. |
I'd like to think Raymond would have approved — or at least argued with me about the spoon.
👉 Sunday Afternoon in Central Park — Art Print
View all my work on Redbubble by searching @kornkobart.
Comments