ChatGPT Fujifilm Recipe | FUJI X WEEKLY

AI Can’t Match the Aesthetic Correctly

*Bike Bags* – Copenhagen, Denmark – Fujifilm X-E5 – Provia Positive

Someone told me, “When I want a Recipe for a certain look, I just ask AI to make it for me.” I’ve heard a similar sentiment a few times. In fact, there are a couple of website and apps that will do it for you. Just upload a picture, and AI will create a Recipe that supposedly matches the picture. ChatGPT and other AI programs can do it, too. But is it any good?

It’s an easy test, actually. Use a photo that was made using a Recipe, and see if it comes up with the same or similar settings. If not, then it doesn’t do a good job. Test it again using a photo made using that same Recipe, and see if it gives you the same or similar settings as the first time. If it doesn’t, then it’s inconsistent and can’t be trusted. In order for AI to be a useful tool for this, it needs to be both accurate and consistent; if neither are true, it’s not a good tool.

If you upload a photo to AI and ask it to create a Fujifilm Recipe, you hope that it’s going to produce reasonably similar results to the photo you uploaded. If not, then what’s the point? You might as well just use random settings, or let darts or dice determine what to program into your camera. Someone once told me that AI is better than a complete novice, but not as good as an expert. I think there’s a lot of truth to that, but when it comes to creating Recipes, AI might actually be worse than the novice.

For example, I uploaded the image at the top, Bike Bags, which was made using my Provia Positive Recipe, and asked ChatGPT to make a Recipe from that photo. You can see my chat with AI below:

I noticed two things immediately. First, it has a lot of similarities to Reggie’s Portra. Certainly not 100% (and for sure it’s not as good), but there some obvious parallels. Second, it says to use base ISO 160 and also DR200; since the Dynamic Range settings are ISO dependent, you can’t use DR200 at ISO 160. A minor grievance, but it illustrates that AI isn’t fully versed in Fujifilm settings. This is how Provia Positive compares to AI’s “Nordic Harbor” (a.k.a. Reggie’s Portra tweak):

The results are pretty divergent. You might like the AI version (it’s basically Reggie’s Portra with a slight green cast), but you can’t say that it’s a good facsimile of the Recipe it’s suppose to look like. Using the same account, I started a new chat and asked ChatGPT again to make a Recipe using the same photo, but I worded the question slightly different. It basically gave the same Recipe but with four changes: Reala Ace instead of Classic Chrome, Color 0 instead of +1, Shadow -2 instead of -1, and WB Shift +1 Red & -3 Blue instead of -2 Red & -4 Blue. Let’s take a look:

ChatGPT’s Reala Ace version of its “Nordic Harbor” Recipe

After this, I used a separate ChatGPT account and repeated step one precisely, and got even different results (this time it suggested Velvia). I then tried using different photos made with the same Provia Positive Recipe, and each time it gave me a different Recipe, and none were very close to my Recipe. Between the two accounts, I asked ChatGPT to make a Recipe based on a Provia Positive photo a total of 10 times. Even if it had given an accurate result once (which it didn’t), how could you trust it with divergent suggestions each time? How could you ever know which one would be most correct?

ChatGPT’s third attempt, this time with Velvia

I then tried other AI programs and apps, and repeated the test. Interestingly enough, on my first attempt with Claude, it gave a Reggie’s Portra like Recipe (just like ChatGPT), but with an even stronger green cast than ChatGPT’s iteration. Gemini also chose Classic Chrome on the first attempt, but played it safe and didn’t venture very far from factory-defaults. After 10 tries with these other AI apps and programs, plus the 10 I did with ChatGPT, I had a total of 20 results. Three things shocked me. First, Classic Chrome was chosen nine times, far more than any other film simulations. I assume this is because Classic Chrome is by far the most popular film sim for those who use Recipes, and that’s influencing AI’s answers. Second, Provia wasn’t picked even once. Not one time did AI choose the correct film simulation—not Provia, the “standard” option. Even on the programs specifically intended for making Fujifilm Recipes from photos, it got it wrong. If I had picked at complete random, most likely I would have chosen Provia at least one time (there are 20 film sims, and I made 20 attempts, but nine of them are monochromatic and not color). Third, one time Acros was given for the film simulation. One of the AI’s thought that Acros—a B&W film simulation—was the most correct! Wow.

This article is already way too long, so let me wrap this up. Yes, AI can make a Recipe that is based on a photo that you upload to it. Anytime you ask AI to do something, it will give you an answer. But is that answer any good? When it comes to Fujifilm Recipes, probably not. First, as you can see above, AI couldn’t figure it out. It also gave a different answer each time, and sometimes very nonsensical ones. Both the inaccuracies and inconsistencies are major red flags, and demonstrate that it’s not a trustworthy source for this particular task. It’ll probably get there someday, but it’s not there today. This is a very easy test, one that you can do yourself. If you ever wonder if any AI program or app can make a Recipe that matches the aesthetic of a picture that you upload to it, simply use an image that you already know the parameters of; if it can’t match it—or at least be in the general ballpark of it—then it’s not a good resource. You might actually be better off throwing darts, rolling dice, or picking the settings out of a hat at random.

Can ChatGPT Replicate a Recipe??

Posted on January 29, 2025 by Ritchie Roesch

Someone pointed out to me that ChatGPT can create a Film Simulation Recipe that replicates the look of any photo that you upload. Just share a picture with AI, and it will analyze the look and produce a Recipe that matches. But can it really do that? And, if so, is it accurate?

So I gave it a try. I uploaded a picture and asked AI to replicate the aesthetic of it with a Fujifilm Recipe. Sure enough, it produced one. I also asked ChatGPT to tell me which Fuji X Weekly Recipes are the closest. Let’s examine the results.

Test 1

The top picture is the camera-made JPEG captured on my Fujifilm GFX100S II using the Nostalgic Film Recipe (a Film Dial Recipe). I uploaded it to ChatGPT and asked it to make a Recipe based on that image. The second picture was reprocessed in-camera using the settings that AI provided. Quite obviously, they are pretty far apart and are not a good match.

Some of the settings are the same. Both pictures use DR400. Both are Grain Weak (although ChatGPT chose Grain size Large). Both have Color Chrome Effect Strong and Color Chrome FX Blue Weak. Noise Reduction is -4 on both. The White Balance Shift is +2 Red & -4 Blue. All of the other settings were different—for example, the Nostalgic Film Recipe calls for the Nostalgic Neg. film simulation, while AI chose Classic Chrome. The Recipe made by AI isn’t too dissimilar to Kodachrome II, although not an exact match. I asked ChatGPT to tell me which Fuji X Weekly Recipes were the closest, and it suggested Kodak Ektar 100 and Kodak Gold 200.

Test 2

The top picture is the camera-made JPEG captured on my Fujifilm GFX100S II using the Fujicolor PRO 160C Warm Recipe (which is also a Film Dial Recipe). The bottom picture was reprocessed in-camera using the settings that ChatGPT provided. They are far apart, and are not an especially good match, although it is slightly better than the first test.

The Fujicolor PRO 160C Warm Recipe uses the Reala Ace film simulation, while ChatGPT chose Eterna. Most of the settings are divergent. I don’t have a Recipe that’s especially close to the one made by ChatGPT, but Timeless Negative is probably the most similar. The Fuji X Weekly Recipes that AI suggested are the closest are CineStill 800T, Kodak Ektachrome 100SW, and Kodak Vision3 250D.

Test 3

The top picture is the bottom image from Test 1. I asked ChatGPT to make a Recipe that matches it, to see if it would suggest the same settings that I used to make it (which is the Recipe AI had made just a few minutes earlier). The Recipe that it came up with was not the same, and also not a great match. Instead of Classic Chrome, it wen’t with Classic Negative. While only somewhat similar, Fujicolor Blue is probably my closest Recipe. The Fuji X Weekly Recipes that AI thought were the closest are Kodak Portra 400 and Kodak Portra 160.

Test 4

I re-uploaded the first photo from Test 1, and asked the same question to see if it would suggest the same settings, or if it would change them. Unsurprisingly, it was completely different, this time choosing Astia for the film simulation (I didn’t bother to reprocess). It suggested that my Kodak Portra 400 Recipe would be the closest match.

Conclusion

I’ve said before that ChatGPT is a neat little card trick, but once you know how it works, it isn’t nearly as impressive. I’ve also said that it will only get better with time. Both of those statements still seem to be true. While AI can analyze a picture and produce some Fujifilm settings to try—which is amazing—it isn’t very accurate. It can also suggest Fuji X Weekly Recipes that might be similar, and it is a little better at that, but still not especially good.

I have zero doubts that with enough time—and it may only be a couple of years—ChatGPT will be able to analyze a photo and produce a Fujifilm Recipe that is pretty similar. It might be even closer to suggesting an existing Recipe that isn’t far off. It’s not there yet, though.

I’m certain that ChatGPT was trained on my website. When asked directly, it says that it wasn’t, but states that it was trained on various photography blogs and forums. When I asked it to name some specific ones, the very first it listed was Fuji X Weekly. So, yeah, AI did some of its learning from my website, which of course I did not consent to nor was I compensated for. I believe that this training is ongoing, and it probably will even learn from this article that you are reading right now. This is wrong, and eventually the courts will likely determine that it is completely illegal; however, by then, the damage is done, and there’s nothing that can be done.

With all of that said, I had a conversation with a friend of mine who is super intelligent (he’s a scientist) about AI and ChatGPT. He said that he is not concerned about being replaced by AI, but he is concerned about being replaced by someone who knows how to use it better than he does. I don’t think it is prudent to ignore ChatGPT completely. While undoubtably ethically sketchy, it is a tool that can sometimes be used effectively, if you understand its strengths and weaknesses. I’m not sure what that means for me specifically, other than it is good to test it now and again to see how far it has or hasn’t come. It definitely has made some progress since the last time I tested, but not a lot.