Surprising results: analyzing more images with ChatGPT

Building on my recent image analysis experiments with ChatGPT, I explored its capabilities more thoroughly. I uploaded numerous images, posed a wide range of questions, and was often pleasantly surprised by the results. In many cases, ChatGPT provided accurate, useful answers. However, some questions led to frustrating dead ends. Still, when it worked, the AI demonstrated impressive skill in analyzing images and offering insightful feedback.

Ask ChatGPT to analyze color saturation and vibrance

Saturation is a dead giveaway for images which have been over processed and pushed too far. We often want colors to be rich, yet natural and authentic. ChatGPT can help analyze saturation levels and tell us whether images are under, over, or realistically saturated. Here's an example prompt:

PROMPT: Analyze the saturation of this image and provide:
1. The average saturation value (0-100 scale).
2. Whether the image is under-saturated, over-saturated, or realistic based on common standards for landscape photography.
3. If saturation is not realistic, specify whether it affects the entire image or only certain hues. If specific hues are affected, identify which ones are most saturated, and provide options for adjusting their saturation using Photoshop.

I tested this prompt with the image below (left), and ChatGPT calculated the average saturation as 54.88 on a scale of 0-100, then followed up with this:

ChatGPT: Based on common standards for landscape photography, this level of saturation is within a realistic range but leans slightly towards being moderately saturated. It is not overly exaggerated, but the warm tones may appear enhanced.

Did you catch the most important part of that response? "Common standards for landscape photography." Standards? What standards? From where?

I asked ChatGPT to clarify where these standards came from and to cite sources if available. The AI responded with a list of sources including photography books, scientific studies on human perception, the Munsell Color System, and digital image analysis of images from "renowned" landscape photographers and National Geographic. I was especially surprised to see a direct reference to National Geographic, so I asked ChatGPT if it had access to the magazine's digital image archive.

ChatGPT: No, I do not have a license to analyze or directly access images from National Geographic or any other copyrighted sources. My reference to National Geographic and professional landscape photography was based on general observations, industry standards, and principles of color science—not on direct image analysis.

So...yeah. Like a lot of things with AI and ChatGPT, it can sometimes be difficult to verify where exactly information is coming from, and whether the AI is inventing its own facts. I pressed ChatGPT for more information, and it created a visual plot comparing my image against the "standard" for saturation.

Fascinating, right? According to ChatGPT, my image has slightly lower saturation than the standard mean (indicated by the red dashed line) and a broader standard deviation (20 versus 15). The latter can be seen in the red tail on the left edge of the chart. What this means is that there's a wider gap between the low-and-highly saturated pixels in my image compared to the norm. I either need to add more vibrance to undersaturated areas, and/or lower the saturation of the most intense hues (red/yellow).

With that, I edited the image to reduce the vibrance of red and yellow (right image above), and the saturation now appears more natural and less luminant. I could probably tweak overall brightness more to match the original, but the colors are more within the range of normalcy.

Ask ChatGPT to analyze and plot skin tones

Video editors have the advantage of using vectorscopes to analyze hues in footage and align skin tones with an industry-standard reference for accuracy. Yet, for some reason, photo editing software has never included this feature. However, we can ask ChatGPT to analyze a portrait, plot its skin tones, and suggest adjustments to refine hue and saturation.

When I uploaded an image of myself, ChatGPT first analyzed it and then asked if I wanted to isolate and plot only the skin tones. In other words, it detected my face and masked just the area I was interested in—pretty cool. Here’s the vectorscope plot it generated:

As expected, my skin tone was pretty much right on the line (because I edited it in DaVinci beforehand), with slight variances around it. But ChatGPT said my skin tone could be further improved:

ChatGPT: The skin tones slightly lean toward cooler tones (higher Cb values, indicating more blue influence). The spread is fairly broad, meaning there is some variation in skin tone across different lighting conditions in the image.

It then followed-up with instructions for DaVinci Resolve and Photoshop to push midtones warmer, and reduce saturation in the red/yellow areas.

I asked ChatGPT if it could verify (prove) it was properly isolating and analyzing my skin tone. Basically, I wanted to know if it was looking at the right thing. It responded with this:

ChatGPT performed reasonably well at isolating and masking my skin, similar to a qualifier in DaVinci, Premiere, or Final Cut. I could also refine the selection further to focus solely on my face, but you get the idea.

Ask ChatGPT if an image is over-sharpened

Everyone worries about sharpening. Too much? Too little? Will viewers think my image looks "crunchy"? Well, we can answer that question by asking ChatGPT to analyze sharpness and provide feedback. To test it, I uploaded the following landscape image of mine.

ChatGPT analyzed the image for sharpness and responded with the following graph and explanation:

ChatGPT: There is a significant amount of edge detail, particularly around rock formations. However, there is no excessive presence of unnatural halos, which suggests that sharpening is within a natural range. [The Laplacian Histogram at right] suggests a natural sharpness distribution without excessive high-frequency edge enhancement. The image is sharp and detailed but does not exhibit strong signs of oversharpening. There are no prominent halos or unnatural contrast boosts that would indicate aggressive sharpening. If you are looking for a softer or more filmic look, you might consider slightly reducing sharpness, but as it stands, the image appears well-balanced.

So, there we go! From a technical perspective, ChatGPT doesn't think I have anything to be worried about. For the record, I over-sharpened another image, uploaded it to ChatGPT, and the AI correctly quantified the over-sharpened edges and detected halos as well.

Ask ChatGPT how a photograph should be printed

I love making prints of my photography, but doing so is a different process than editing images for screen. On paper, images are no longer backlit, so they can often appear darker than expected. Paper type also plays a strong role in how much contrast is present and how bright the whites and highlights are. It is an art unto itself.

I asked ChatGPT to help prepare this black and white image for printing with the following prompt.

PROMPT: Analyze this landscape image and tell me:
1) Which paper would be most appropriate for the subject
2) What edits I should make to improve the image when printed on fine art paper
3) Whether additional sharpening should be applied prior to print

To my surprise, ChatGPT didn't simply recommend paper type (matte, glossy, etc), but specific brands of paper, including Hahnemühle Photo Rag Baryta (ironically, one of my favorites), Canson Baryta Prestige, and Hahnemühle Photo Rag Matte. Plus the following:

ChatGPT: If you want the richest blacks and a more contemporary feel, go with Baryta papers. If you prefer a more subtle, painterly rendering, opt for matte cotton rag.

Spot on!

As for edits, ChatGPT recommended boosting midtone contrast to enhance detail, selectively deepening the blacks, plus a slight recovery to the highlights to keep bright values from venturing too far into white (and losing detail). For sharpening, ChatGPT was again spot on — apply sharpening at the output stage after the image has been resized for print.

I then decided to push ChatGPT a little further and see if it would recommend the same papers with a similar, yet very different color image:

Here again, ChatGPT surprised me with the following:

ChatGPT: This abstract aerial landscape features soft gradients, rich blues, and intricate textures. To preserve the depth and color vibrancy, the best paper choices would be:

Canson Infinity Platine Fibre Rag (310gsm) – A semi-gloss cotton paper that enhances deep shadows while maintaining fine detail.
Hahnemühle Photo Rag Metallic (340gsm) – If you want to enhance the shimmering, fluid quality of the blues, the metallic sheen will give it an ethereal glow.
Hahnemühle FineArt Baryta Satin (300gsm) – A great choice if you want deep blacks and a smooth transition in tonal gradations.
Epson Hot Press Bright (Smooth Matte, 330gsm) – If you prefer a softer, more painterly look with less contrast but excellent shadow detail.

First, I never mentioned to ChatGPT that this was an aerial landscape photo! The AI could have easily mistaken it for a macro shot of a tiny subject, yet it correctly identified the image’s scale. Second, I was particularly surprised by its recommendation to use metallic paper—a choice I usually don’t prefer. However, after considering its analysis, I can see why metallic might be a compelling option for this type of image.

Ask ChatGPT to find locations

ChatGPT is surprisingly adept at analyzing images and guessing their location. Almost every time I've asked the AI to guess where an image was captured, it gets it right. We can put this knowledge to practical use—especially for landscape photography—by asking ChatGPT to provide recommendations for similar landscapes to explore and photograph. For example:

PROMPT: Guess the location of this image, then share five similar landscapes around the world I can visit, with Google Maps links to each if available.

I uploaded a photo taken along the Oregon coast in the United States, and ChatGPT accurately identified its location. It then suggested several similar destinations worldwide, including Durdle Door (England), Wharariki Beach (New Zealand), Playa de las Catedrales (Spain), El Arco de Cabo San Lucas (Mexico), and La Portada (Chile).

If I were curating a collection of images based on a specific theme or concept, this kind of insight would be incredibly valuable. Additionally, you can specify places you’d prefer to avoid—such as heavily photographed national parks or over-touristed locations—to ensure a more unique and original location.

Ask ChatGPT to select and sequence portfolio images

When grouping images together as part of a photography project, the order of images can be just as important as which images are included. ChatGPT is not a professional, seasoned photo editor by any means, but it can shuffle the deck and formulate a reasonable framework by which to organize images.

I tried this by uploading 10 images from a recent landscape photography trip to White Sands, along with this prompt:

PROMPT: Analyze and sequence these images to create a pleasing, engaging narrative flow from start to finish. Afterwards, explain why the sequence works, why each image was positioned where you placed it, and whether any images should be excluded.

ChatGPT analyzed my images and grouped them by "Foundational" (Core Narrative & Visual Impact), "Supporting" (Bridges Between Key Moments), and "Possible Exclusions". For the latter, it picked an image it said was "somewhat redundant in terms of composition and tone."

From there, it created a recommended sequence:

CHATGPT: The sequencing should create a natural rhythm and evoke a journey through light and space. Here is my proposed flow:
Opening image — Image 7 (A soft, open dunescape with a muted sky, inviting the viewer in)
Image 1 (A gradual introduction of form and gentle color)
Image 3 (A slight shift in angle and texture, building intrigue)
Image 2 (The dunes become more defined, with richer light)
[...and so on]

ChatGPT then asked if I wanted a mockup (!) of my images in this sequence. I replied "Yes", but the AI failed to create one (too intensive, perhaps). Nonetheless, the textual order it described was sufficient for me to experiment with the sequencing for myself (which wasn't half bad!)

Tips for better answers from ChatGPT

As mentioned in my earlier exploration of photography and ChatGPT, I believe the AI is mostly analyzing images against tried-and-true best practices and regurgitating information from countless blog posts, books, videos, etc. So I don't take its advice all that seriously. However, it does often surprise me with helpful information that actually makes a qualitative difference when editing images.

If you're interested in exploring ChatGPT's analytical capabilities with your own photography, I recommend the following:

1. Be explicit with your query — Be direct and blunt. You're speaking with a bot, not a human being, so there's no need to be polite! Tell ChatGPT exactly what you're looking for and how it can be most helpful.
2. Upload low resolution images — In my experience, ChatGPT can sometimes struggle when analyzing images and return annoying server errors. As a workaround, output a small, medium-quality JPG of an image prior to upload (eg, 800px on its longest edge and 50 quality).
3. Call its bluff — ChatGPT may be smart, but it can be wrong. If the AI shares information that doesn't pass the smell test, ask it for proof. Ask it how it arrived at an answer, what logic it used, quantify its answer (if applicable), and provide sources to backup its claims. This can lead to surprising and in sometimes illuminating insights into its process.
4. Repeat image analysis — In addition to be wrong, ChatGPT can sometimes contradict itself. For example, I processed the same image with the same prompt using the same model (GPT-4o) and received completely different results. I pointed out this discrepancy, and the AI stated it "may have used an incorrect formula" in my first request. Lesson learned, don't accept what the AI says on face value.
5. Provide context and examples — Tell ChatGPT where to look online for examples, or provide it with images to learn from. The more context and information you can provide, the better the result.
6. Assign a persona — It may feel weird, but you can get better answers by telling ChatGPT who it is. For example, "You are an impressionist painter", "...cinematographer", "...photographer", "...fine art photography printer", etc. Begin your query by placing ChatGPT in the right headspace, then ask your question.
7. Ask follow-up questions — If ChatGPT references a word, acronym, or concept you don't fully understand, ask it for more information! Treat it like the world's largest and smartest dictionary.