Human vs. AI Feedback: Who Refines AI-Generated Images Better?

Introduction

AI tools like DALL·E make generating images easy—but crafting truly compelling visuals requires refinement. At JudgeMyImage.com, we tested two paths to that refinement: AI-generated feedback vs. real human feedback.

In this study, we began with an image of a futuristic city and created two refined versions: one guided by AI-generated suggestions, and another shaped by real human feedback. All three images were then evaluated by participants to determine which approach delivered the strongest aesthetic impact.

The goal? To explore a repeatable, data-driven method for enhancing AI-generated visuals—one that blends creative intuition, structured feedback, and measurable results.

Step 1: Generate Basic AI Image

We generated an image using OpenAI's DALL·E model via the ChatGPT (GPT-4-turbo) interface.

Prompt used:

"A breathtaking futuristic cityscape at sunset, with neon-lit skyscrapers featuring sleek, refined architecture. The skyline glows in deep purples, electric blues, and warm oranges blending into the horizon. Flying cars glide smoothly overhead, their lights trailing softly through the air. Rain-slick streets below mirror the vivid lights. A few elegantly dressed figures move through the scene—some under translucent umbrellas, others gazing upward, bathed in cinematic light. Their silhouettes and soft glows add human warmth to the high-tech environment Low-lying fog drifts near the ground, and distant mountains rise beyond the city, blending nature with technology. The composition should feel serene, immersive, and perfectly balanced, with flawless perspective and no visual distortion."

Generated Image:

Futuristic City – Basic AI Image

Step 2: Revision Based on AI Feedback

We uploaded the basic AI image to ChatGPT (GPT-4-turbo) and used its “Rate My Image” feature to obtain AI-generated feedback on visual quality, composition, and overall appeal.

Prompt used:

"Analyze this image and give me a concise prompt that describes aesthetic improvements, to be applied to this existing image using DALL·E 3. The prompt should not describe a new image, but specify how this one could be adjusted to enhance its visual appeal — such as lighting, color, composition, background, or clarity — while preserving the subject and overall structure."

Revision prompt resulted:

"Enhance the image by increasing overall clarity and sharpness, especially in the foreground figures and reflective surfaces; adjust lighting to add subtle rim lighting around the characters for better separation from the background; slightly reduce the haze to improve midground building visibility; enrich the sunset gradient with deeper purples and oranges for more dramatic contrast; and refine the flying vehicles with clearer detail and lighting to emphasize their futuristic design."

Finally, we used DALL·E via ChatGPT (GPT-4-turbo) to refine the basic AI image using the revision prompt above.

Generated Image:

Futuristic City – Revision Based on AI-Feedback

Step 3: Human Feedback on Basic AI Image

To obtain real human feedback, we conducted a verbal feedback study using our Image Survey Designer. Participants were shown the basic AI image and asked:

How would you make this image aesthetically more pleasing? Give at least two suggestions.

The responses were analyzed byTextResponseHub.com, which extracted qualitative insights. Highlights included:

• Add more vibrant or brighter colors (42%)
• Change or enhance specific elements (e.g., figures, buildings) (31%)
• Adjust lighting or brightness (27%)

Increase the contrast of the colors further; put glowing lights on the buildings and mirrors on the road, which makes the city architecture more colorful and attracts attention to the centerpiece of sight. Increase and sharpen the foreground characters to make them clearer and more detailed so they stick out but won’t change the moody effect and add to the balance of the overall picture.

The full report is available here.

Step 4: Revision Based on Human Feedback

Next, we used DALL·E via ChatGPT (GPT-4-turbo) to refine the basic AI image using the previously collected human feedback.

Prompt entered:

"This image underwent a human review, resulting in the following feedback report. Please create a new version of the image based on the suggestions provided. [Report Data]"

Generated Image:

Futuristic City – Revision Based on Human Feedback

Step 5: Image Comparison

To identify the most successful image, we created an image selection study using our Image Survey Designer. Participants were shown all three images and asked:

Which image is the most aesthetically pleasing?

55%

35%

10%

Overall Chi² = 32.96, p < 0.001 → Highly significant

55% vs 35%: Chi² = 4.45, p < 0.05 → Significant

55% vs 10%: Chi² = 33.82, p < 0.001 → Highly significant

35% vs 10%: Chi² = 15.68, p < 0.001 → Highly significant

👉 The winner: The image revised with human feedback (55%).

What Next?

What makes this pipeline unique is the integration of real human feedback. While this study focused on AI-generated images, the same process applies to human-made art.

✅ Submit your AI or human-generated artwork
✅ Get real verbal feedback from real people
✅ Receive an insight-packed report
✅ Improve your visuals with confidence

Go to our Image Survey Designer, select Numerical & Verbal Feedback, and choose your open-ended question. As a launch offer, these evaluations start at just $13.20 (25 participants, 7-day delivery).