The Treachery of Images

How I Created a Luxury Brand Ad with $1.20 (and Why That's Interesting)

You don’t know, but you do know that this image isn’t real.

You know this because I tell you it’s fake. Because I include the legal disclaimer text that essentially says “don’t trust your eyes!”. I’m creating a feeling of dissonance - like a classic Magritte - only more concerning since this isn’t just a painting of a cigar on a wall.

This image took me roughly two hours to generate as a riff off of one of Creative Director Saint Louvent’s 30 ads in 30 days instagram exercises. I was fascinated by how compelling these AI works were and was curious about the process. So like any good artist and technologist, I tried to recreate one of her images to learn.

For my image, I used a combination of Midjourney (v6.1, $8 per month), Claude.ai ($0), and Procreate ($12.99 purchase). I am not an advertiser. I am not a graphic designer. I have zero experience in marketing or branding and limited experience with the image generation side (GenAI) of the LLM world. I do however have a strong visual inclination and an orientation towards data, tech, art, good value, and a disdain for gatekeeping.

If I prorate creating this based on tooling costs this comes out to about $1.20 in materials. That’s both totally stunning and also completely absurd on its face – but I’ll address that absurdity in another post.

I found this whole process of creating and then looking at my creation fascinating and wanted to share the process and the reflection on the process and the product.

My initial thoughts and observations:

1. System 1 Thinking defaults will absolutely steer viewers toward any established brand architecture when sufficiently activated, even if inaccurately so. The colors, items, font choices, framing, and spacing directly play Tiffany’s “brand story” – one they’ve invested in for decades.
   
   Therefore, even though I have not selected the precise fonts or the typical grammar or ad copy, I have selected enough referential content to convey that this is part of the Tiffany and Co. The choices that mattered were being in the zone with the Tiffany blue, a serif font, and silverware.
   
   Learn more about System 1 and System 2 thinking here (TL;DR)
   Learn more in depth from in Thinking Fast and Slow by Daniel Kahneman here.
   

2. Simple concepts and minimalistic ideas are easier to produce than accurate replications of brand designs like patterned logos for example (versus shapes) because we have a stronger sense of “what is correct”. Life drawing is considered one of the most difficult genres of art because humans are so habituated to 'looking at people' that deviations from the norm (e.g., proportions, stance, etc.) are very obvious. The same principle applies to LLMs.

   
   I want to distinguish here between text-based LLMs and tasks that can be measured in a classic machine learning ala precision and recall or F-score (e.g. extract for me the latest analyst stock price prediction for LVMUY).
   
   Precision solves for - did it provide the right answer?

   Recall solves for - did it provide all instances of the right answer?
   Learn more about calculating these metrics here.
   
   

3. Many images can be precisely rendered without being accurately rendered. Similarly, it is difficult to consistently render (although this is getting better). Details and non-precise representational art (e.g., surrealism, abstract art, etc.) can act as a form of creativity and also as a means of visual misdirection from technical limitations. Avoid trying to force accurate details via GenAI at this stage, it’s much more powerful to blend with a tool like Photoshop.
   
   For example, the recent AI music video drop from Ye is very visually detailed and rich. Note that the choices don’t have an obvious, generally familiar, “accurate” reference (human faces). There’s a lot of object / character distortion and the more traditional human faces are hidden or obstructed.
   
   This can be both a powerful stylistic and storytelling choice and also a convenient choice given the current state of model capabilities. (I am not promoting Ye only noting that this video is very visually interesting and whoever created it or prompted it into being is very skilled).
   
   

4. However, visual accuracy can be simulated by incorporating other tools as a post-processing step. Text, logos, etc. are easier to manipulate with a Photoshop-like tool than it is directly via a model. The model is unlikely to generate any commercial quality content without some kind of additional tool integration, even if it is just to add ad copy.
   

5. GenAI does not obey, it takes suggestions. Oh, and it doesn’t care about grammar. Write in tag format and lean on common turns of phrase and concepts. Locate well known visual references (e.g. Murakami or Annie Leibovitz) and think of the output as a visual search mashup. It does help to be well-versed in specific visual references and/or techniques (e.g. 35mm) that have likely been well-documented online. You are guessing at its underlying training data - and honestly, it’s not hard to land a hit. These models are enormous and you can likely assume there were some liberties taken with copyright (ethics be damned! *sarcasm*). But practically, if lot of people know about it, it’s probably capable of being referenced.

At the start of my musing, I specifically observed the role that my “AI disclaimer” played in making me feel a bit more uncomfortable. It’s essentially saying HEY WAIT - THIS IS FAKE. It says wait because I’v placed it in the bottom left corner for your eyes to process this after you take in the image holistically. I’m making an assumption that your eyes (if you’re reading this in English) scan content from left to right and top to bottom. Therefore I’ve strategically placed the text in order to activate System 1 first.

The discomfort, I believe, comes from a dual conflict: first, between accepting something as potentially real while knowing it isn't, and second, from being jolted from comfortable System 1 thinking into analytical System 2 thinking. This forced transition compels me to re-examine the image, now actively searching for signs of artificiality

Besides serving as token play on Magritte, the disclaimer is an actual CYA 1) because I do think it’s unethical to not label it but also 2) because I’m referencing a real brand and therefore reaching into trademark territory if I don’t denote that this is essentially a “Spec Ad”.

Again, I was not hired by Tiffany and this is not real. If you’re interested in the prompt itself please send me a note! This brief exercise was super fun to do but obviously the fun conceals an entire industrial complex that is GenAI and LLMs, ethics, copyright, climate impact, etc. which I’ll continue to explore on this stack.

See safe out there!