What was asked, what was done
What the owner said
I want to generate several images and have the AI pick the best one, without spending my subscription usage.
So here is what I did
- I showed six knight images one by one and scored them out of 10.
- I re-scored them using a 100-point scale with instructions to list flaws first.
- I combined the images into a single numbered set and asked it to pick 'best' / 'worst'.
- I shuffled the order five times to verify if it selected the same image.
What the owner didDefining the selection criteria in a single sentence
Initial approach: Scoring one by one out of 10 failed
I showed six knight images one by one and scored them out of 10. The result was that all six received a score of 10. Since they were tied, the 'first place' was simply treated as the first image.
Switching to a 100-point scale + "list flaws first" yielded similar results. The most awkward image (the one with tangled arms) actually received high scores of 85, 82, and 87. Scores for the same image also fluctuated between 78 and 87.
Revised approach: Picking from a single numbered set
I combined up to six images into a single numbered sheet. Then, I instructed it to pick 'best' / 'worst' in a single line of JSON.
The answer format I asked for was one line of JSON: {"가장좋음":번호,"가장나쁨":번호,"이유":"…"} (the keys are Korean for best, worst and reason; 번호 means the sheet number).
I shuffled the image order and ran multiple trials to tally the scores. 'Best' received +1, and 'worst' received -1.
It selected the same image even after shuffling the order five times
I tested this by shuffling six knight images five times (6 images x 5 trials = 10.0 seconds). The result matched human judgment.
| Item | Result | My Judgment |
|---|---|---|
| Best | First image created with klein (5/5 trials) | Highest quality ✔ |
| Worst | Image with tangled arms (5/5 trials) | Arms are tangled and the pose is awkward ✔ |
I changed the criteria to "thick outlines, flat color 2D cartoon style" and ran three trials with eight images. The 3D-style image was selected as 'worst' in all 3/3 trials. This indicates it is reading the criteria.
Limitations
It is good at filtering out bad ones but weak at ranking among good ones. In the test, four images were tied. It missed subtle flaws (the image with an abnormal cloak tail).
The test was conducted using one type of image and one person's judgment as a reference. Since identical inputs yield identical answers (temperature 0), the effect of multiple trials only comes from shuffling the order. The scores are for ranking, not absolute quality metrics.