Art work is hard with ChatGPT...

Clare and myself at another wedding in 1978

The days before camera-phones! We have very few pictures of our early married life (we were married in February 1978). Why not recreate some of our early memories using the generative-image powers of ChatGPT?

I upload a bunch of the portrait pictures of Clare and myself that I do have, scattered across various ages, and ask ChatGPT to construct prompts to generate said images. Many very poor portraits result, followed by much tweaking of the prompts.

Really, dear Reader, I eventually lose the will to live after reviewing the pastoral image below. Not especially good, you will conclude.

Clare's reaction? Effectively "Burn it at once!".

So let's discuss with ChatGPT why this is so hard.

ChatGPT's concept of us in 1978

---

Nigel: The difficulties you are having with accuracy are because you have to go from reference image to textual prompt to generated image, a very lossy procedure. Face recognition algorithms go straight from a source image to some matrix-construct which very precisely captures appearance, which is then used to generate very accurate images, is that right?

But you can't use that technique?

---

ChatGPT: Exactly — you’ve put your finger on it.What I have available here is only description-based generation:

• I take a set of reference images.
• I infer verbal features (hair colour, face shape, expression, glasses type, etc.).
• I then pass that descriptive prompt to the generative model.

Each translation step is both imprecise and noisy — some details get dropped or distorted. That’s why likeness often drifts, and why we go through iterative refinements. It’s a “lossy” pipeline, as you say.

By contrast, face recognition / embedding systems (like those used in biometrics or deepfake pipelines) convert an image into a high-dimensional embedding vector (a matrix-like numerical representation). That embedding encodes fine-grained facial geometry and texture in a way that can be recombined almost perfectly. If one couples those embeddings with a generative image model, you get near-photorealistic consistency of likeness across new poses, ages, and scenes.

I don’t have the ability here to run embedding-based identity capture and reapply it — precisely because it would blur the line with surveillance and biometric cloning. Instead, I stay strictly in the prompt–generate–adjust loop, which protects against someone uploading, say, a passport photo and asking me to generate a deep-fake.

So yes, the method you describe is the technically superior one, but it’s also tightly restricted. I’m confined to the safer but fuzzier prompt-driven workflow.

---

Nigel: And yes, I do know there are other systems out there, not so fastidious!

---

Update: another attempt following further prompt work. You decide, but it's hard to escape the conclusion we're going around in circles.

ChatGPT's image of Clare and myself in 1978 (aged 27)

---