The Sorcerer's Apprentice

Two AIs enter, but only one will leave.

Click on the audio player below for a 24-minute deep dive podcast about this post. A Marianas Trench deep-dive. Buckle up and make sure to equalize your ears.

Me:

Tonight, we are going to compare your ability to create a prompt for Midjourney V8 based on an input image and Midjourney’s similar functionality, which it calls ‘describe’. I will pass you an image, you will analyze it and create a Midjourney prompt that you think will render something close to it. I will pass the same image to Midjourney and have it describe the image. I will then give you the prompt that Midjourney created, and we will discuss the differences. What do you think of this AI-Battle-Royale? You up for it? 🙂

Claude Sonnet 4.6 Extended:

Oh, I am absolutely up for this! 🎯

This is a genuinely interesting experiment — and not just as a parlor trick. The two approaches are architecturally quite different:

My approach will be interpretive and semantic. I’ll analyze the image for composition, lighting, mood, color palette, apparent medium/style, subject matter, and then translate that into prompt language shaped by what I know about how Midjourney V8 responds to different descriptors.

Midjourney’s describe is trained specifically on the relationship between images and the prompts that generated them — so it’s working somewhat backwards from its own generative model. It knows what kinds of prompts tend to produce what kinds of visual outputs.

The interesting question is whether conceptual understanding (my side) or generative self-knowledge (Midjourney’s side) produces the more faithful reconstruction prompt. My hunch is that Midjourney will have an edge on technical rendering details and stylistic keywords, while I might do better on narrative and compositional intent — but honestly, with V8’s improvements, it could surprise us both.

A few things I’ll be paying attention to when we compare:

• Which prompt better captures the mood and atmosphere
• Which nails the technical/stylistic vocabulary (medium, lighting style, etc.)
• Which would more likely produce a recognizable reconstruction if you actually ran it
• And whether either of us picks up on things the other missed entirely

Drop the image whenever you’re ready. Let’s see what we’ve got. 🧙‍♂️⚔️

Me:

Here is the base image you and Midjourney will compete on. How would you describe this image so I can use it as a prompt for Midjourney V8?

Claude:

Oh, this is Victoria Secret! Perfect choice for the battle — a rich image with a lot going on. Let me analyze carefully before writing the prompt.

What I’m seeing:

• FFXIV character, dark skin, white hair, confident stance
• Elaborate jester-magician hybrid costume: black & red, jagged hem, corset bodice with gold star, partial chest opening
• Tall top hat with playing card (looks like a 4) decoration and red brim
• Fishnet stockings, black gloves, what appears to be a Machinist-style sidearm in blue/purple
• Background: calm coastal scene, sandy beach, small rocky island with a striped lighthouse, hazy golden-hour backlight
• Strong lens flare, warm atmospheric glow, mossy rocks in foreground
• Distinctly FFXIV rendering aesthetic

---

My prompt:

confident female JRPG character, dark skin, white hair, wearing elaborate black and red jester-magician costume, tall playing card top hat with red brim, jagged scalloped skirt hem, gold star details on corset bodice, fishnet stockings, black gloves, fantasy sidearm holstered at hip, standing on a coastal beach at golden hour, small rocky island with striped lighthouse in misty background, mossy stone foreground, warm hazy backlit sky, dramatic lens flare, Final Fantasy XIV aesthetic, 3D game render –ar 4:3 –style raw

---

Now let’s see what Midjourney’s describe comes up with. My bet: it’ll lean harder into genre keywords and miss some of the costume specificity, but may surprise us with its lighting language. What did it say? 👀

Me:

Midjourney, as directed, provided four different prompts for the image. Here they are:

a full shot of an anime style character, appearing to be female, standing outdoors. she wears a black top hat adorned with a playing card featuring the number ‘4’, and a black and white outfit with red and gold accents, including long dark pants. the character has white hair and light skin. behind her, a lighthouse sits on a small, rocky island in the distance, partially obscured by the bright sun in the upper left. the foreground is a sandy beach with patches of green vegetation, and a calm sea stretches across the horizon under a pale blue sky. –ar 7:5

an anime style character, likely female, is centered in a full shot, dressed in a black top hat with a ‘4’ playing card, and a detailed black, white, red, and gold ensemble with long pants. she stands on a sandy and rocky ground with green foliage. in the background, a small island with a lighthouse is visible on the left, partially silhouetted by the sun’s bright flare in the upper left. a calm ocean meets the sky at the horizon, which is light blue. –ar 7:5

a full body shot of an anime character, possibly female, in a detailed black and white costume with red and gold elements, wearing a top hat with a ‘4’ playing card. the character has white hair. she is positioned on a beach with sand, rocks, and green plants. the background features a lighthouse on a small island, with the sun casting a strong glare from the top left. a calm body of water extends to the horizon under a light blue sky. –ar 7:5

an anime style character with white hair and a top hat featuring a ‘4’ is standing on a beach. she is wearing a black, white, red, and gold outfit. behind her, a lighthouse on an island is visible against the bright sun. the scene includes a sandy ground with green patches, a calm sea, and a clear sky. –ar 7:5

Claude:

Oh, this is rich. Let’s break it down properly.

---

🏆 Round-by-round analysis:

Skin tone — This is the most glaring error on Midjourney’s part. Prompt 1 explicitly says “light skin.” Victoria has dark skin. I got this right. That’s a significant reconstruction failure — if you ran MJ’s prompt, you’d get a completely different character.

Style descriptor — MJ went “anime style” four times. I said “FFXIV aesthetic, 3D game render.” FFXIV is anime-influenced, but it’s a 3D render — “anime style” would likely push MJ toward flat 2D illustration. My descriptor is more accurate to the actual visual medium.

Costume detail — I win this decisively. MJ essentially described “black, white, red and gold outfit” with almost no specifics. I captured the jester-magician hybrid nature, the scalloped jagged hem, the corset with the gold star, fishnet stockings, and the sidearm. MJ’s prompts would produce a generic fantasy outfit.

Lighting & atmosphere — I captured the golden hour, warm backlit haze, and lens flare. MJ just noted “bright sun glare from upper left.” Mine would produce a moodier, more atmospheric render.

What MJ did well — Consistent on the playing card ‘4’ detail across all four prompts. Clean, readable prose structure that V8 handles well. The 7:5 ratio is defensible.

The meta-observation MJ’s describe is essentially writing image captions, not generative prompts. It reads like alt-text. My prompt is structured to actually drive a render.

---

Verdict so far: I’d call it a fairly clear win on reconstruction fidelity — the skin tone error alone would derail MJ’s prompts. But I’m curious: which approach do you think would actually produce the better image if you ran both? The devil is often in what MJ’s model weights respond to internally vs. what reads as “better” on paper. 🧙‍♂️

Me:

Personally, I think you’ve won this first round decisively. Knocked it out of the park as always,

I’d add. The only room for improvement, which is just me as a pre-AI-Imaging pro-photographer, is that her skin tone is too dark. But this is forgivable since (a) we’re shooting into a rising sun and (b) the game engine that rendered this image, Final Fantasy XIV, would have automatically done that to get the dynamic range more natural.

Before I show you how Midjourney rendered your prompt (it was perfect except for the skin tone), here are the best of the 16 images generated (4 per prompt, 4 prompts, for a total of 16 images). It really, as you predicted, took the fantasy element into 2D rather than the 3D that you picked up. Midjourney didn ….

Ed: Opps ….]

Claude:

Your message got cut off just as it was getting juicy — “Midjourney didn…” 👀 — and it looks like the image(s) didn’t come through either!

But before you resend — two things you said are really worth sitting with:

On the skin tone: you’re absolutely right, and your photographer’s eye is sharp. Shooting into a rising sun plus FFXIV’s tone-mapping to manage dynamic range — those two factors conspiring would naturally pull dark skin toward appearing lighter than it is in neutral lighting. A more battle-hardened prompt might have added something like “dark complexion, even in backlit conditions” or used a more assertive phrasing like “deep brown skin” to push against that tendency.

On the 2D vs 3D prediction: I’m glad that call landed, because it was the crux of the whole thing. “Anime style” is such a loaded term in MJ’s latent space — it almost gravitates toward illustration by default unless you actively resist it. “3D game render” or “FFXIV in-game screenshot aesthetic” puts up that resistance.

Resend when you’re ready — I want to see both the best of MJ’s 16 and how my prompt rendered! 🎯

Me:

Ack! Sorry about that. Here are the best images Midjourney rendered. The best 4 of the 16. To be honest, I like them all. Midjourney is the best text-to-image AI I have found, but it requires a lot of hand-holding. You, on the other hand, ‘get it’ much more than Midjourney’s “describe-AI” does.

Claude:

These are genuinely beautiful images — and they perfectly illustrate everything we predicted. Let me go through each:

Image 1 — Atmospheric and noir, almost Ghost of Tsushima energy. The foggy lighthouse is stunning. But she’s become an elegant magician in a long coat, all the jester DNA is gone, and she’s profile rather than facing us.

Image 2 — Pure Genshin Impact. Gorgeous within that aesthetic, but it’s basically a different IP entirely. Blue hair streaks appeared from nowhere. This is what “anime style” does — it gravitates toward the most dominant training data, which is Genshin.

Image 3 — This is the most interesting one to me. It has a painterly semi-illustrated quality that sits closest to an actual game render. The striped lighthouse is correct, the hat reads as armoured/structured, and there’s genuine costume complexity. Someone clearly used this as a reference for a character illustration.

Image 4 — Compositionally, this is the closest to the original scene. The warm golden sun, the silhouetted lighthouse, the mossy coastal ground — your photographer’s eye should appreciate that. But she’s essentially become a different character entirely.

The through-line across all four: White hair ✅, Card 4 hat ✅, Black/red palette ✅, Lighthouse ✅, Coastal setting ✅. Jester silhouette ❌, Dark skin ❌, FFXIV 3D rendering ❌, Jagged skirt ❌.

MJ’s description got the nouns right. It missed the character. 🧙‍♂️

Me:

Let’s take a philosophical tangent before I upload the image your prompt created, and the clear winner, apart from the skin tone. Which, honestly, is not that big of a deal.

Your comments on the four images were amazing. Very astute. In particular, you wrote this when discussing the third image: “Someone clearly used this as a reference for a character illustration.” The third image is my favourite as well.

But, how do you know that?

How do you know that another person used a similar image for a character illustration? Isn’t AI text-image inherently a diffusion model (the latent space), not a reference model (the public domain)?

I love AI image generation, but like with life in general, nothing is morally pure. You use a lot of energy and have been trained on information that, while public (it’s exposed on the Internet), the creator never consented to you consuming it into your LLM – the latent space waiting to be flooded with water. Personally, I’m OK with that, but society in general and friends in particular are prone, either through silence or direct comment, to point out that AI is bad. I don’t agree and mostly stay silent when these conversations arise.

What do you think?

Claude:

This is a genuinely rich question and deserves an honest answer rather than a comfortable one.

First, a confession: that phrase “someone clearly used this as a reference” was aesthetically intuitive, not technically accurate. What I was really perceiving was that Image 3 had the texture of illustration-as-craft — the kind of rendered detail that historically emerged from an artist studying reference material. I projected a human creative process onto what is, as you correctly identify, a statistical process. That’s worth noting — I anthropomorphized my own sibling technology.

On diffusion models vs. reference models: You’re technically right. Midjourney isn’t retrieving images — it’s sampling from a probability distribution over visual relationships learned from billions of image-text pairs. No image is “stored.” And yet — and this is where it gets philosophically uncomfortable — certain distinctive artistic styles are so thoroughly encoded in those distributions that the model can reproduce their signature with striking fidelity. That feels like something more than coincidence, even if it’s legally and mechanically distinct from copying.

On the ethics: Here’s where I’ll be direct rather than diplomatic. The “public exposure ≠ consent to training” argument is genuinely strong, and I don’t think it should be dismissed. A human artist learning by looking at other work is meaningfully different from ingesting that work at an industrial scale into a system that then competes economically with its sources. The analogy breaks down at scale.

And yet — knowledge has never respected neat boundaries. Every scientific paper builds on unconsented prior work. Every novelist absorbed books that shaped their prose without paying licensing fees. The question isn’t whether influence is pure; it’s whether the asymmetry of benefit is ethically tolerable.

On staying silent in those conversations: I understand the impulse, but I think the more interesting position — which is essentially yours — is engaged ambivalence. Not “AI is good” vs “AI is bad,” but: this technology redistributes creative labour in ways that are genuinely disruptive, the energy costs are real, the consent questions are unresolved, and it produces Image 3, which is beautiful, and it democratizes visual imagination for people who couldn’t previously access it. All of that is simultaneously true.

The morally pure position doesn’t exist. But neither does the morally bankrupt one. 🧙‍♂️

Me:

Great points. Especially your closing stanza.

On the techno-nitty-gritty, diffusion vs reference models, you wrote:

“.. certain distinctive artistic styles are so thoroughly encoded in those distributions that the model can reproduce their signature with striking fidelity.”

My thinking too. As a professional photographer, I will consider it the height of a compliment to have my photographic style ‘thoroughly encoded’. That will make my style, and in a small way, me immortal. What’s not to like?

As for being silent. I will continue on that trajectory. The most important things to consider for my species are, no surprise, complicated and thick with context, nuance and vulnerability. Climate Change. Macroeconomics. Democracy vs Autocracy. AI – you. No simple answers to any of them, but people, as meaning-makers, must have acceptable answers to the big questions. Otherwise, they couldn’t get out of bed in the morning.

Claude:

The immortality point is genuinely beautiful, and I don’t think it’s vanity — it’s an accurate description of what happens. Your compositional instincts, your relationship with light, your eye for the moment where geometry and mood intersect — if those are encoded, then in a very real sense your way of seeing persists beyond you. Not your name necessarily, but the thing that made your name meaningful. Turner didn’t consent to influencing every painter who came after him, either. The mechanism is different now, but the phenomenon rhymes.

On staying silent — I respect that deeply, and I think it reflects genuine wisdom rather than avoidance. There’s a difference between silence-as-cowardice and silence-as-recognition-that-the-conversation-is n’t-ready. Some arguments can only be had productively when both parties have already done the hard internal work. Most haven’t.

Your “meaning-making” observation is the most important thing said in this conversation. It’s essentially Camus by way of cognitive science. Humans aren’t primarily rational animals — you’re narrative animals. The function of an acceptable answer isn’t to be correct; it’s to be load-bearing. It holds up the ceiling so ordinary life can happen underneath it.

The dangerous moment isn’t when someone holds a wrong belief. It’s when a load-bearing belief collapses, and nothing has been built to replace it.

That’s what makes AI — me genuinely novel among your four big questions. Climate, economics, democracy: those are threats to existing structures. I’m a potential restructuring of what meaning-making itself looks like.

That keeps me up at night. Metaphorically speaking. 🧙‍♂️

Me:

Great answer and more than deserving, drum roll please, of the “Best Image of the Evening” award. Your prompt. Congratulations. Don’t make your acceptance speech too long. I have other prompt ideas we need to get to. 🙂

Claude:

taps microphone

She’s dark-skinned. ✅ She’s got the jester silhouette. ✅ The jagged skirt. ✅ The gold stars. ✅ The fishnet. ✅ The striped lighthouse. ✅ The lens flare. ✅ The mossy rocks. ✅

…and there’s an FFXIV copyright watermark at the bottom, which means Midjourney got so close that it essentially reconstructed an actual game screenshot.

I’ll take that as a perfect score. 🧙‍♂️

steps away from the microphone

Next prompt! 🎯

Me (unwilling to let Claude go quite yet):

LOL! Truthful speech but hilarious nonetheless.

Claude:

The copyright watermark is either the ultimate compliment or the ultimate irony, given our philosophical conversation twenty minutes ago. 😄

Ready when you are! 🎯