> ## Content Index
> Fetch the complete content index at: https://mitchellens.ink/llms.txt
> Use this file to discover other available public pages before exploring further.

# Closer #013
- URL: https://mitchellens.ink/closer/013/
- Published: 2025-07-02T18:00:09.000Z
- Updated: 2025-10-06T14:33:13.000Z
- Description: I'm reconsidering my approach, but I'm not sure I'm willing to change it. Not right now. These are still interesting times we live in.
- Author: Mitchel Lensink
- Tags: Closer, Field

SPONSORED 

This weekly experiment is on an indefinite hold, until significant AI-advancements have been made.  
  
Curious about the conclusions? Read [Photography in the Age of AI (2025): Lessons From 25 Weeks of Experimentation](https://mitchellens.ink/closer/photography-in-the-age-of-ai/)

[Read now ](https://mitchellens.ink/closer/photography-in-the-age-of-ai/) 

Welcome to Closer,

A weekly experiment in AI interpretation: I share a photo, ask AI to describe it, then generate an image from that description. This is edition #013\. Prompts are in grey boxes. *Descriptions are in italics*.

Small changes to the prompt again this week! Should I continue making small changes, or do you want me to make bigger ones? I feel like there's a balance to strike between doing wildly different things and being consistent.

Either way, the end goal is still to investigate the quality of AI image generation over time. Over the years, probably even. That's why I try to keep these as concise as I can, including the descriptions AI is generating for us. Although I'm breaking that rule a little this week. Let's start.

## LLMs used

**GPT-4o** for image interpretation and image generation.

## The input

✍️

Please describe the attached image as well as you can.

![The image shared with ChatGPT as input.](https://storage.ghost.io/c/8a/15/8a1564aa-a008-4e21-8398-12d04de07fcb/content/images/2025/06/010_Mitchel-Lensink_220804_United-Kingdom-London-1.jpg)

The image shared with ChatGPT as input.

## The output

*The image shows a close-up of a wrought iron fence with a geometric, repeating circular pattern. Nestled between the bars, caught gently within the iron curves, is a single red rose. The rose is wilted—its petals are drooping and slightly darkened, suggesting it has been there for some time. The green leaves and stem are still attached, hanging down through the lower gap in the fence.*

*The metal of the fence is weathered and dark, with a matte texture that contrasts softly with the organic delicacy of the flower. The background is slightly blurred but indicates an urban setting, likely a city street or a public space. The interplay of hard metal and fading flower evokes a quiet, melancholic atmosphere—perhaps a tribute, a forgotten gesture, or simply something left behind.*

*The photo compositionally balances symmetry and emotion: the repetitive, cold geometry of the fence is broken only by the warm, fading red of the rose. The image invites reflection and subtly suggests themes of memory, loss, and unnoticed beauty in the everyday.*

✍️

Generate an image with an aspect ratio of 2:3 based on this description: {**the description previously generated*}

![The image generated by ChatGPT based on its description of my image alone.](https://storage.ghost.io/c/8a/15/8a1564aa-a008-4e21-8398-12d04de07fcb/content/images/2025/06/ChatGPT-Image-Jun-30--2025-at-11_00_22-PM-Large.jpeg)

The image generated by ChatGPT based on its description of my image alone.

## Impressions

Some observations:

- Making the prompt more elaborate leads to more elaborate descriptions. Which leads to more accurate image generation.
- This week we went back to the roots of the original prompt we used at the beginning of this experiment. I left out most of the nuance after the first sentence though. It turns out: that *does* matter.
- Which means, in this case, asking the AI to describe something 'as well as you can' does not give us 'as well as it can'. Not without those further instructions beyond that.

Some questions:

- How are most people prompting their image generation?
- What does the average person put into the interface to quickly make an image?
- Are most people simply saying 'give me an image' or are most saying 'give me an image and think about this, do that, consider this too, don't incorporate those things, stay away from this and that'?
- More elementary to this newsletter: which approach are we trying to test consistently here?
- Should we keep the consistency, to be able to accurately compare the outputs over time to see if the technology gets better as we progress, or should we aim for the best possible output we can get?
- What is more fun?

What I think about the description and the generated image:

- the description of the image is more about the emotions it evokes and less about what physical elements are visible. I think that's interesting, because the emotions conveyed by an image are very subjective.
- What's continuously interesting is that when you are reading the AI-generated description with the original image in your head, you can totally picture it. It makes sense. Then you see the AI-output and you break free from the illusion that there's only one version of the reality described.
- Of course I'm heavily biased but I greatly prefer my image from the AI-generated one. The stylistic choices made for that image are... very mediocre. Like your dad, who has a job in IT and recently bought a Canon DSLR with a 24-135mm lens, is suddenly trying to be 'artsy'. Cringe.

I don't think we can conclude the technology is not significantly advancing, based on the lacklustre images coming from our simple prompts to get the descriptions. It's a bit hit or miss so far with the output. I'm having fun though. Plus I think the *real* change will come when new LLM models will release over time.

I think it's good that we're building this library of tests with the current tools available. It sets a baseline. Something to refer back to, the moment these tools become more and more normalized. It makes me wish we started this newsletter three years ago, when the outputs still looked objectively, hilariously, bad. I think we passed *that* stage but I also think human creators needn't worry just yet. Perhaps we never need to worry, but that will depend on the kind of work you make too. Those thoughts I still have to sit with a little longer. Perhaps run some more experiments. Are you still with me?

Replies are welcome, as always! Please let me know if you have questions, suggestions or remarks about my approach. This is just as much a journey for me as it is for you.

See you next week.

Mitch