Can you use Artificial Intelligence to create Film Simulation Recipes for Fujifilm cameras? Does AI even know what that is? If so, would those Recipes be any good? Those questions and more have been rattling around inside my head for the last few months.
Back in January I asked Open AI‘s Chat GPT what a Film Simulation Recipe was and it didn’t know—it couldn’t differentiate a Film Simulation Recipe from a Film Simulation, but only went so far as to acknowledge that Film Sims could be customized. It also didn’t know much about Fuji X Weekly. It didn’t take long for things to change.
Last week Open AI announced the new-and-improved GPT-4, and so I put it through the same test. To my surprise, it not only knew what a Film Simulation Recipe is, but could even create one! The AI also provided the reasons why it chose the settings it did, and they seemed logical. However, I noticed that some required parameters were missing, so I asked it to add those settings to the Recipe, and it did. I also asked Chat GPT to give the Recipe a name. The AI created Recipe turned out to be pretty good, and I used it for some photography in southern Arizona, but I’m getting ahead of myself. We’ll get back to this in a moment, but let’s talk about some other things first.
I wanted to find out how Chat GPT knew how to create a Film Simulation Recipe. Obviously the software has never used a camera, so where was it getting its information? I asked, and the software deflected, telling me where I could find further information on Fujifilm cameras and Film Simulation Recipes. To my surprise, Fuji X Weekly was listed as a potential resource. I wasn’t satisfied with the answer, so I kept asking, rewording the question, until the software admitted that it gathers information from (among other things) websites, such as Fuji X Weekly, One Camera One Lens, Dan Bailey, and Ted Forbes. I was a little surprised on the last two, especially Ted who I’m pretty certain doesn’t use Recipes or out-of-camera JPEGs. Maybe it just threw out some names associated with Fujifilm gear. My conclusion is that the software searches the web for resources and attempts to make sense of what it finds, and it borrows heavily from the work of others (but, thankfully, doesn’t outright copy).
I had Chat GPT make a total of seven Film Simulation Recipes for the Fujifilm X100V. Not once did it provide me with all the necessary parameters on the first try, and I had to ask it to add the missing fields. It always chose Auto White Balance (once it did not provide a WB, so I had to ask it to do so), and only provided a White Balance Shift once on the first try; for the other six Recipes, when I asked it to provide me with a WB Shift, it never ventured beyond +/- 2 for Red and Blue. In fact, only once did the software suggest that any setting go beyond +/- 2, so I think it takes a conservative approach, perhaps not understanding why anyone would want to go wild with the JPEG options.
On the first try I was provided with Color Chrome Effect and Color Chrome FX Blue settings: Off on both. For the next five attempts Chat GPT didn’t list either option, so I had to ask, and it chose Off each time. With the final Recipe attempt, it listed Color Chrome Effect set to Strong, but didn’t list Color Chrome FX Blue; when I asked it to give me a Color Chrome FX Blue setting, it (surprisingly) stated that Color Chrome FX Blue isn’t an option on the Fujifilm X100V. I think the software struggles to understand what these settings do and why someone would choose them, and also struggles to understand what specific settings are available on each Fujifilm model.
Two settings that the software never provided on the first try are Grain size and Clarity. It would list Grain strength (Off, Weak, or Strong) but never size (Small or Large). When asked about size, it suggested Large four times, Small twice, and Standard once (there is no Standard option). With Clarity, it typically suggested a positive number, and only gave me a negative number once, zero once, and Off (which I suppose is the same as zero) once. As with the other settings, it never ventured beyond +/- 2.
I asked it to mimic the look of a certain film stock, and Chat GPT provided a Film Simulation Recipe that (in my opinion) wasn’t a great match. I then asked it to mimic the look of a different film that just so happens to be the same exact emulsion just sold under a different brand name (to see if it would provide similar or identical settings), and the second Recipe was much different than the first. It’s clear that the AI isn’t analyzing pictures from film to create its Recipes, but instead finds descriptions of the stocks and suggests which Fujifilm settings could logically match the descriptions. “Vibrant” means Velvia and “soft” means PRO Neg. Std, which makes sense to a point; even though one film can produce many different aesthetics based on how it was shot, developed, printed and/or scanned (among other things), I believe you’d be hard pressed to find a single emulsion that could be emulated by both Velvia and PRO Neg. Std, but that’s what the software did.
I also asked Chat GPT to create a Recipe that I already have a Film Simulation Recipe for. I wanted to see if it would just copy my Recipe, but thankfully it didn’t. It was actually significantly different. For now, at least, the software isn’t outright plagiarizing anyone (that I’m aware of), and I hope it stays that way. Finally, using a different account, I asked Chat GPT to create a Recipe with an identical request to one of the seven, just to see if it would give me the same answer, and it didn’t. I repeated this test once more, and it once again provided a different result. While it tries to come across as “intelligence” it appears to be more like a roll of the dice.
Based on this test, I believe that AI is about 60% of the way there to being a useful tool for creating Film Simulation Recipes. It struggles to know which parameters to provide. It doesn’t understand the nuances between camera models. It seems to take a rather predictable and conservative approach to creating Recipes. While I think it tries to be logical with its choices, it is basically just taking a guess and giving random settings, which might produce good results sometimes and might not other times. The amazing thing, though, is that just a couple months ago the AI didn’t even know what a Film Simulation Recipe was, so it has made significant strides in a short period. I think eventually—and it might not even be that far out—the software will be able to analyze an image and provide settings for your Fujifilm camera that will be a reasonably close match to that image. For now, though, AI isn’t a particularly good way to get a Recipe.
I promised that I would get back to the first Film Simulation Recipe that I asked Chat GPT to create for me to use on my Fujifilm X100V. I asked the software to create a Recipe that would be good for an urban environment at night. It provided me with everything except for Grain size, White Balance Shift, and Clarity, so I asked it to give me those settings, too. Then I asked it to name the Recipe—it gave me five choices, and I went with the first: Urban Dreams.
I found Urban Dreams to be a pretty decent and versatile Recipe. I didn’t get a chance to use it much in an urban environment at night, but the results were good in the few opportunities that I did have. It seems to do well in sunny daylight, dreary overcast, nighttime, indoors, outdoors, landscapes, street, still-life, portraits, etc., so this could be one’s go-to Recipe for everyday photography. It reminds me a little of Kodachrome 200, a high-ISO slide film introduced by Kodak in the mid-1970’s and discontinued in 2006. It’s not completely “right” for Kodachrome 200, but can be surprisingly close sometimes, and this Recipe is probably the closest to it that I’ve seen.
This Urban Dreams Film Simulation Recipe is compatible with most X-Trans IV cameras: the X-Pro3, X100V, X-T4, X-S10, X-E4, and X-T30 II. For X-Trans V and newer GFX, you can use this Recipe, but it will render slightly different (try it anyway). For the X-T3, X-T30, X-Trans III, and older GFX, simply ignore Clarity and consider Sharpness set to 0; the results will be very similar (only slightly different).
Dynamic Range: DR400
Noise Reduction: -2
Grain Effect: Strong, Large
Color Chrome Effect: Off
Color Chrome Effect Blue: Off
White Balance: Auto, 0 Red & +2 Blue
ISO: Auto, up to ISO 6400
Exposure Compensation: +1/3 to +1 (typically)
Example photographs, all camera-made JPEGs using this Urban Dreams Film Simulation Recipe on a Fujifilm X100V:
This post contains affiliate links, and if you make a purchase using my links I’ll be compensated a small amount for it.
Find this Film Simulation Recipe and nearly 300 more on the Fuji X Weekly — Film Recipes App! Consider becoming a Patron subscriber to unlock the best App experience and to support Fuji X Weekly.
Help Fuji X Weekly
Nobody pays me to write the content found on fujixweekly.com. There’s a real cost to operating and maintaining this site, not to mention all the time that I pour into it. If you appreciated this article, please consider making a one-time gift contribution. Thank you!