Althea case study

The elephant in the room: if I never received the samples, then anyone, anywhere in the world, could take a project like this from start to end, at any price.

As Claude would say, fair hit, and it's been top of mind as I've explored AI more and more. My experience at the time of this project's completion, January 2026, was that it wasn't easy, and there were compromises when this project was aimed to be entirely AI generated. But the pipeline I built meant the final imagery isn't easily distinguishable from real photography, i.e. it isn't slop.

At the time, that pipeline was still experimental in stages. Althea was my fifth AI project. The four before it had each tried a different way of getting to final assets I felt were photoreal, and far enough away from AI slop.

The brief

Althea was a client job, with a budget and timeline behind it. Knowing what images I could deliver, without ending up in a situation where I'd overpromised, was the most important thing when attempting to deliver a project with unproven production tools.

I approached it as a pre-production-heavy project. I always spent significant time in pre-production, but the non-deterministic nature of AI models being used as the output component added layers of complexity. Many of the questions I'd usually ask, about art direction, the shoot location and the logistics of moving samples, weren't relevant for this shoot. It was much more about which lane, of the infinite lanes AI can be used for, Althea should be defined by.

My client's moodboards had all the important information. Our first meeting focused on the rebrand being built around a clinical approach, rather than the naturalistic scenes products like this usually sit in.

So: clean lines, crisp hard light, and clean white as the backdrop and dominant visual theme. The colours could then carry the mood of each variant without being visually loud.

Below are examples of the brand deck moodboard and product images I was given. They're mockups, but after a detailed discussion I was able to note the kind of plastic and the scale of each product.

Early testing

Before the moodboard and final packaging mockups arrived, I spent some exploratory time testing how I could control composition and developed a 3D pipeline, which I wrote about in Unlocking the AI magic box with 3D blockouts.

Those early experiments were about one question: could I generate these in a time-efficient way?

Glass and dolerite

Once the moodboard and product images were supplied, my testing revolved around distilling the most context-rich themes of the Tasmanian landscape, and finding man-made materials with a clarity that signalled clinical and clean.

Glass was the natural choice. Its visual complexity was a good contrast to the packaging designs, and to the brand's home in Tasmania. I explored every kind of glass, finding myself circling back to gradient glass. It had visual impact, but it was restrained enough to be used at different depths depending on the image. In some it could be a lighter shade, in others the glass could be thinner. Composition and intended impact decided how each was used.

The other key theme was stone, in this case dolerite. The challenge was how to bring in a clean interpretation of the surface without it becoming the main element in the scene.

In early testing I merged the two: a machined stone with a gradient glass cap. Results were something not commonly seen, and echoed some of the products' purpose: calm, and the phase shift between states. The stacked, chipped glass also had natural wavy forms, so it read like waves or water.

I also tested a stacked version, the two combined as one piece. It reminded me of resin-filled stone. But I wanted something lighter, with a storytelling element that belonged to this brand.

A head tilt to look at it another way

The idea of presenting a mountain peak without the natural terrain evolved into machined stone, in a colourless white, invoking a clinical look without removing the core visual identity of the mountain. The sides laser-cut smooth and the top left natural created a contrast which made the glass blend more convincing. Moving to a more powdery, pure white also reduced grain variance and shadows. The piece kept its visual complexity without the weight, and let me align the glass in a precise, engineered, purposeful way.

After I presented the first ideas, the final iteration included an option with a flat pane of glass hovering over the mountain. The client chose this approach, which moved the project into the next phase. Generating a night-time scene was simpler with these geometry rules.

It also removed a lot of ambiguity about how the products would rest on surfaces too. It was now just flat glass, so building mockups for the image models was much easier. Every scene with that surface had stable objects. If I needed to give the model a finished image 1 to show scale and placement on the glass for image 2, image 1 could influence the prompt much more than objects at different heights and angles.

Making these decisions early, and not showing or even testing ideas I now knew would be very hard to replicate at scale and on time with AI, paid off in the second half of the project. I could focus on the bigger problems that would deliver far more interesting imagery.

Scale and the logo problem

Scale across different colours of the same container was relatively easy. Scale across different objects was very difficult. And scale was one of the most important components that couldn't be allowed to drift.

I ran early tests with scale-correct 3D blockouts, but that pipeline was too inefficient for the time and budget available for testing new ideas and scene angles on these hero images. So I worked with a two-pass process instead. First, generate the scene with every element except the products, locking composition, lighting and effects. Then use that base and lock the model with heavy prompting and clear visual references of the refined product images.

A core issue I knew I'd have was the logo. The model could follow the Althea logo's general shape, but there's an important detail on the capital A that it failed on constantly. And the smaller the text and the more objects in a scene, the more the variant names drifted.

This is where my 90% rule came in. Rather than keep re-rolling for those elements, I generated until the products' scale and scene integration was photoreal, then inpainted the logo and text where needed. If all failed, which it did, I retouched the logo manually. The text is fully fixed in the stills, although, as expected, the video drifts heavily, and that's still true with the latest models.

Next time - late 2026 and beyond

Models have improved significantly since the start of 2026 when this project was completed. There's every chance this pipeline, with complete logos, might be a single-pass prompt by now. But for future projects where logos and branding have to be tack sharp and accurate, I'd still always generate the product images to a clean, crisp standard first. That includes matching exposure, colour and even texture across every product image.

And if you're wondering: yes, it can mean half a day editing an image that the AI only needed one minute to generate.

See the finished project →