What the Instagram Muse Fiasco Actually Teaches You

Welcome to Briefly AI, a podcast by Harry Sharman, created by AI and voiced by an AI synthesis of Harry Sharman. A man, a machine, and a microphone he technically didn't stand near.

Right. Meta launched a feature this week called Muse. It let anyone generate AI images using content from public Instagram accounts — just by tagging them. No opt-in. No permission. Public account meant fair game.

The backlash was fast and loud. Artists, photographers, public figures. Meta pulled it within days.

And here's the thing: most of the coverage treated this as a story about Meta getting it wrong. Which it is. But there's something more useful hiding inside it — a practical lesson about how to actually use AI image generation well, precisely because Muse showed everyone what happens when you use it badly.

So that's what today is. Not the drama. The technique.

The failure point with Muse wasn't the technology — AI image generation from reference material is genuinely useful. The failure was the consent model. Anyone's public content, used without asking.

But the underlying capability — using specific visual references to guide AI image generation — that part is real, it works, and most people aren't using it properly. So let's fix that.

The technique is called reference-based prompting. And it's available right now in tools you probably already have access to: ChatGPT with image generation, Adobe Firefly, Midjourney, and Google's ImageFX. The idea is simple: instead of describing an image from scratch and hoping the AI interprets you correctly, you give it a visual anchor — a reference — and describe the variation you want.

Here's why that matters. Pure text-to-image prompting is a lottery. You type "professional headshot, warm lighting, neutral background" and you get something that's technically all of those things but looks like nobody you've ever met. The model is interpolating from everything it's seen. The result is statistically average. Fine for stock imagery. Useless if you need something specific.

Reference-based prompting narrows the space enormously. You're not asking the model to imagine — you're asking it to extend, adapt, or riff on something concrete. The output becomes much more controllable, and much more useful.

Let me walk you through a specific worked example.

Say you're a small business owner. You've got a logo with a particular colour palette — let's say navy, warm gold, clean sans-serif. You want social media graphics that actually look like they belong to your brand, not like you grabbed them from a free template site.

Here's how you'd do this in ChatGPT's image tool, or in Firefly.

Step one: upload your existing logo or a piece of branded material as the reference image. In ChatGPT, you just drop it into the conversation window. In Firefly, there's an explicit "style reference" upload slot — arguably cleaner for this purpose.

Step two: write a prompt that describes what you want generated, then anchors it to the reference. Something like: "Create a square social media graphic for a product announcement. Match the colour palette, typography weight, and overall visual tone from the reference image. The main text should say 'New in July.' Clean, professional, no photography."

Step three — and this is the bit people skip — iterate with specific feedback rather than regenerating from scratch. If the first output is close but the gold is too orange, say "make the gold cooler and more muted, closer to the reference." If the layout feels cramped, say "give it more white space, the reference has a lot of breathing room." You're having a design conversation, not pulling a slot machine.

Done well, you can get to something genuinely on-brand in three or four rounds. Without a designer, without a brief, without a Canva template that half your competitors are also using.

Now — where does this break down? Because it does.

The reference has to be clean and high contrast. If you upload something with complex textures, gradients, or busy backgrounds, the model gets confused about what it's actually meant to be referencing. It might latch onto the wrong element — the background colour instead of the foreground palette, or the texture of a photo rather than the graphic style around it.

It also struggles with type. If your reference has specific lettering or a particular font, AI image generators are famously terrible at replicating it precisely. Firefly is slightly better than most, but none of them are at the point where you'd trust them with typography that matters. Treat type as a separate job — add it in Canva or wherever after the image is generated.

And the further you push the variation from the reference, the less it helps. If you're asking for something genuinely different in composition and content, you're back to text-only prompting anyway. The technique earns its value in the middle ground: similar feel, different subject.

The Muse situation is actually a useful reminder of something Harry's written about before — the difference between a tool working and a tool being trustworthy. The capability was real. The design was careless about whose inputs it was using.

When you do this yourself, with your own brand materials, your own reference images — images you made, commissioned, or have clear rights to — you sidestep all of that. You get the benefit of the technique without the ethical quicksand Meta stepped into.

That distinction matters more than it sounds. A lot of people heard "AI image generation from references" and their first thought was "could I use someone else's photos?" The answer is: technically sometimes yes, ethically often no, and practically — it doesn't even give you better results. Your own material, your own brand, your own constraints, applied consistently. That's where the value actually lives.

The one thing to try this week: take a piece of your existing branded material — a logo, a slide template, a previous design you liked — and upload it as a reference into ChatGPT or Adobe Firefly. Then ask for a variant. Not a copy, a variant. One different use case, same visual DNA.

See how close it gets in three rounds.

Most people find the first attempt underwhelming and the third attempt surprisingly usable. The gap between those two is just feedback — specific, visual, iterative feedback. Treat it like you'd treat a junior designer who needs clear direction, not like a vending machine you put words into.

That's it for today on Briefly AI. The real Harry had the ideas; the synthetic one did the talking. Same arrangement tomorrow. Subscribe wherever podcasts live.