The a16z Show Quietly Made the Case for Deterministic AI Image Generation in Email

I listened to the a16z Show episode “Building AI for Creators | Luma & Phota Labs” expecting a conversation about prettier pictures. I got something more useful: a sharper way to think about AI image generation in email. Yoko Li sits down with Luma’s Matt Tancik and Phota Labs’ Zach Xia, two people building at the front edge of generative imaging, and the spine of the discussion is not “the model can make anything.” It is the opposite. The interesting work now is control.

They keep circling one idea: why the future of creative tools may depend less on generating content and more on helping people express ideas they couldn’t easily realize before. Read that with a marketer’s brain. The win is not infinite output. The win is a precise, intended result you can trust, produced the same way every time. Generation got cheap. Getting the exact thing you wanted, reliably, is the part that still costs you.

That one distinction decides how AI image generation in email should work, and most brands are betting on the wrong side of it. The popular pitch is to let a model paint a fresh image for every subscriber at the moment of open. It demos beautifully. At scale it is expensive, non-reproducible, and one bad render away from an embarrassing brand moment. The architecture that holds up flips the order: AI makes reusable source assets upstream, a human signs off, and a deterministic engine composes the final pixel-perfect image from locked layers and live data. Power moves up the workflow. Control stays at the end, where the customer actually sees it.

What the a16z Show Got Right About Creative AI

Tancik and Xia are not AI skeptics. They build models for a living, and they are clear that today’s systems can produce almost anything you ask for. Their point is about where the hard, defensible work sits now. When a model can spin up a thousand variations in a second, the variation is not the prize. The prize is steering it to the one output you actually intended, then reproducing that on demand.

This is a quiet reversal of the last two years of hype. The pitch used to be raw capability: look what it can dream up. The guests describe a more grown-up phase where creators want a specific result, a particular composition, a brand look, an exact subject, and they want the tool to get them there without a hundred rerolls. Generation is the commodity. Controllability is the moat. Hold that thought, because email is where it gets tested with real money on the line.

The Same Shift Is Coming for AI Image Generation in Email

Translate “a precise intended output” into a marketer’s requirements and you get a strict list. The real product has to render exactly. The price has to be the right price. The customer’s name has to be spelled correctly. The offer and the expiry have to match what your commerce system actually says. And all of it has to look identical for ten subscribers or ten million. That is brand fidelity, and it is non-negotiable in a channel where the image carries the message.

There are three places AI can sit in an email workflow, and it helps to name them. First, deciding what to show: which product, which template, which variant. Second, making source assets: backgrounds, lifestyle scenes, AI-extended product imagery for your asset library. Third, producing the final personalized image each individual person opens. AI earns its keep in the first two. The argument over AI image generation in email is really an argument about that third slot, and it is the one slot where a probabilistic model should not be holding the pen. For the math behind why, this breakdown of why generation breaks at scale is the cleanest version I have seen.

Why Generative-at-Open-Time Breaks: Cost, Reproducibility, Brand Fidelity

Start with cost, because it ends most of these debates fast. GPT Image-1 runs around $0.167 per image at the API level. Round it to roughly $0.20 once you add the overhead of doing this in production. A one-million-subscriber program that personalizes imagery across its sends can produce on the order of 1.46 billion images a year. Multiply it out and you are near $292 million annually, just to render pictures, before latency or quality control enter the conversation. Those figures are Zembula’s internal modeling, not a third-party benchmark, but the order of magnitude is the point. Generation will never pencil out at true personalization scale.

Reproducibility is the second problem. A generative model is non-deterministic by design. The same inputs can produce a different image on two different opens, which means you cannot guarantee what any given customer sees, cannot reliably A/B test what you served, and cannot reproduce a render when something looks wrong. You lose the audit trail at the exact moment you need it.

Then there is fidelity, the one that actually gets a brand in trouble. We wrote about REI’s two-handlebar AI ad, a bike rendered with a malformed handlebar that shipped to the public. A model that is confidently wrong does not warn you. Text makes it worse: generative models still mangle words, and in email your words are a price, a name, an offer. Copy correctness is brand fidelity. On top of that, email cannot even be trusted to display your fonts. Litmus has documented for years that only a handful of typefaces render reliably across clients, and that Gmail strips external font imports server-side and falls back to Arial or Times New Roman. The only way to guarantee your brand typeface is to bake it into the rendered image.

This stops being abstract when you look at what acquisition now costs. Average ecommerce ROAS fell to 2.87 in 2025, and ecommerce customer acquisition costs have climbed roughly 40% over two years. When paid media gets this expensive, the owned channel has to carry more revenue, and you cannot afford to put an unpredictable model in front of an audience you already paid to earn. If you want that comparison spelled out, our 2025 email performance benchmark report lays out where email stacks up against paid.

The Architecture That Works for AI Image Generation in Email

Here is the build that respects the a16z principle. Use generative AI exactly where it shines: upstream, to create source assets once. A campaign background. A studio shot with its background cleanly extended. A seasonal scene for the asset library. These are made deliberately, reviewed, and stored. Then the final personalized image a subscriber opens is composed deterministically from brand-locked layers, live data, and explicit rules. Same inputs, same pixels, every single time. That is the version of AI image generation in email that survives contact with a million-person send.

A composition engine is the mechanism that does this. It assembles the final image at request time from layered components you control, so the product is the real product, the type is your type, and the price comes from your data instead of a model’s imagination. One dynamic image URL can render hundreds of use cases without a new template, which is how the personalization economics actually work.

This is the governance answer too. The question is never whether to use AI, it is where in the workflow you put it. Our framework on what you hand to the machine makes the same call: AI at the right stage compounds value, AI at the final render is where it costs you.

The Human QA Gate Before Anything Reaches the Library

Generative source assets do not get a free pass into production. The discipline that makes this safe is a two-step gate. AI generates assets in bulk, an automated QA pass flags fidelity problems, then a human reviews the flagged set and spot-checks the rest. Only approved assets enter the asset library. Nothing a model produced reaches a customer without a person having said yes.

The test that gate enforces is simple and absolute. Would the real product, and any model wearing or holding it, render exactly and unmodified? Background removal, cropping, extending a neutral studio backdrop, upscaling: in the outputs I have examined, these are generally safe when the product itself comes through untouched. Relighting a complex lifestyle scene or quietly reshaping the product is not, and I am watching those closely. Any alteration to the actual item for sale is not personalization. It is false advertising, and no efficiency gain is worth that line.

What Controllable AI Imagery Unlocks, and How to Measure It

Get the architecture right and the upside is real. McKinsey’s research found that faster-growing companies drive 40 percent more of their revenue from personalization than slower-growing peers, with typical revenue lift in the 10 to 15 percent range. A brand could apply that to the broadcast emails it sends today that carry no personalized imagery at all, and the recovered revenue per subscriber adds up quickly. I am keeping that hypothetical on purpose, because the only honest way to claim a number is to measure your own.

That measurement is the part performance marketers will care about most. When the final image is deterministic, you can attribute it. Block-level revenue per thousand and click-to-conversion let you measure a personalized image the way you measure a paid placement, then test and optimize it on the same terms. That is the whole case for treating email as a performance marketing channel: owned audience, first-party data, and measurement you control. Generative-at-open-time cannot give you any of it, because you cannot test what you cannot reproduce. So the right way to read the a16z conversation is as a green light, not a warning. AI image generation in email belongs in your stack. It just belongs upstream, making the raw material, while the engine that composes the final pixel stays deterministic and accountable.

Key takeaways

• Control is the moat, not generation. The a16z guests argue creative AI’s value has moved from making content to realizing a precise intended output. AI image generation in email proves it.
• Name the bucket. AI belongs in decisioning and source-asset creation. The final personalized image each customer sees belongs to a deterministic engine.
• The cost math is brutal. At roughly $0.20 per image across 1.46 billion annual images, a one-million-subscriber program would spend near $292 million a year generating final images (Zembula internal modeling). Composition avoids it.
• Fidelity is non-negotiable. The real product, price, name, and offer must render exactly, every time. Bake the brand typeface into the image so Gmail cannot swap it for Arial.
• Gate every generated asset. AI QA flags errors, a human approves, and only then does an asset enter the library.
• Measure it like media. Deterministic rendering lets you attribute block-level revenue and click-to-conversion, which is what makes email a performance channel you can optimize against your paid spend.