Unlocking the AI magic box with 3D blockouts

I found out very quickly when I started generating images in mid-2025 that I am not very good at describing a scene in detail. How unnatural is it to try and create something visual, using a dictionary?

That sounds like a human prompting problem: someone who lacks the experience to communicate complex visual ideas with words and maybe a few reference images. I found writing a paragraph about a living room that was modern, minimal, with a moss green colour theme and gold accents, always generated something impressive. However, run the same prompt again and you get something like your last, but the sofa has moved, and now has 7 fingers instead of the normal 6.

After many days and weeks, it was apparent from my initial testing that the description was never the constraint. The model has been trained on millions, even billions of images of every living room, and the goal was: how can I get it to generate that one living room from its training data?

Thinking about control

So I started small, with still life. I found a great image from The Ordinary and wanted to see if I could recreate the scene and lighting in 3D, with a placeholder object at centre, then swap out the product for my own. The idea was to build the simplest possible scene — lighting and materials already set — and have the model replace only the subject, working from my reference on white.

The exploration was open-ended, with the end result being a smaller, tabletop-sized scene, and a pathway for me if I needed control.

It worked, and at the time, I had put no thought into whether this scaled. I did conclude at that point that the front-end time spent building a 3D model for the model meant it wasn't commercially viable in that state.

The magic of AI is that it is a shortcut through the entire creative pipeline, leaving only the thinking and judgement to the human. But I was intent on learning and building my own robust workflows rather than following prompt guides. This early progress guided me to the next question that started to form as I tested further: how little do I actually have to build?

A single room

When I started work on what would become Xeriscape, I had the benefit of time testing simpler scenes, and enough evidence by then in this pipeline, and wanted to know whether a model would hold coherence not just for a simple still life set, but for an entire room — whether it could carry geometry, furniture and lighting, and then extend the scene outward.

Xeriscape started as a simple outpaint test, something I knew AI models were able to do from my research. I tested if I could take a simple photograph I had shot against a plain white wall in studio, and then outpaint it. That was the entire task. It worked, but settling on the final room became never-ending in the options. After dozens of versions, all of which were great, I finally landed on a scene – a Palm Springs/desert-themed home.

An entire home

When I first attempted the kitchen as the first wholly AI next room of this home, I started with the completed bedroom, which was created with inpainting and a text prompt. I used a simple 2D floorplan I made, with an arrow, to attempt to convey scale and spatial context. That generated strong results. But the composition changed too much between generations, and I found that describing each component and its place in an image meant that the more words I spent trying to put the stools in place, the more the model somehow weighted the stools as most important, and the rest of the scene elements would drift.

Blender - a program I love dearly.

💡My lightbulb moment was whether I could relearn the basics of Blender and control both geometry and lighting: build basic 3D scenes, then have the model make minimal geometry changes, but the most significant material and lighting changes. To measure all other lanes in comparison - this interior doesn't exist, so you can’t photograph it, and with 3D, it would take a significant amount of time (although not longer than the time I spent on this project), and the result would have the typical 3D-rendered look. So this is an entirely new lane that did not exist only a few years ago, and every step forward, even when taken with steps sideways or back, felt significant.

If I was able to overcome the endless failed generations and testing, it felt like a level of control that would make it possible to generate entirely made-up spaces, or work within client-commissioned built spaces, with AI deployed strategically for the hardest part of the pipeline: materials, textures and fine detail, from 3D to photoreal. After testing models before and after, there is only one on the market currently that can use a 3D blockout/scene as a strict scene and geometry lock. That was Nano Banana Pro in November 2025, and almost a year later, it is the only model that is able to inpaint at this level of precision, at the cost of fidelity when compared to the latest models at the time of this post – October 2026. From my testing, there is always drift, so the decision becomes: do I continue to work on my 3D model, to make sure it doesn't drift from the dining table if I design the structure as intended, and only have it add material and light?

Also, until mid-this year, it was the only model capable of generating images at a production-level resolution, 4K, with photorealistic results. Its earlier model, Nano Banana, was only 1K and significantly less intelligent.

This started an experimentation phase that lasted, on and off, over months while I was working on adjacent AI projects.

Something I discovered across hundreds of generations, with dozens at a time on a single image, is that the first image is simply magic. To most humans, if the model was able to generate this incredible image, which fools most people with its photorealism, then naturally a second, side view of this same space should be easy. I found out quickly this was the hardest part, and the most often misunderstood about AI image models. I found my challenge.

Refinement phase

I have tried many methods to align a model to do something so simple to a human. I have a one-point perspective image of a kitchen. I want to move the camera back to show the living room attached. I have the kitchen, and I have the living room right wall. What worked went beyond marking coloured guide boxes — it was carefully planned coloured numbers and letters, used to condition the model with highly unambiguous information, with text prompting steering how those were read into the blockout scene.

For further workflow context beyond imagery or motion: a large part of my pipeline is having a chatbot format and pressure-test my prompts as the complexity grows. For something dense with multiple reference images, where the same geometry and furniture appear from different angles, I first upload the images to Claude with no context and ask it to describe them. If it can, I can usually hand off and spend less time explaining. If it can't, I use an incognito chat with the same model to strip any context window drift from prior chats it might be referencing. Formatting and removing my typos is the last step, not the most important one — because when sending a prompt to an image or video model, there is no way to pre-test it before committing credits. When working on this pipeline in late 2025, computer vision, which is what chatbots have as a part of their toolset, was still simplistic; the latest models, Opus 4.1 > 5.5, have significantly more grounded vision capabilities that drift less.

Landscape late 2026

While building and refining my 3D pipeline since late 2025, I felt this level of control would eventually be integrated into platforms, and models would be trained to better understand prompts like this. For all the control I tried to carefully apply to my generations, the results had far too much variance for it to ever be viable in traditional pipelines, where iteration and timelines for deliverables were the most important factors.

And only by having Claude run a web search on whether this article is something that is common knowledge did it return the current landscape of prompting. Some platforms are now integrating native workflows like this, with the ability to generate complex 3D scenes in browser, with no 3D knowledge, and then text prompt the finishes on top.

So where does 2027 go? Agents are making the complex more accessible. The scale matters. For simple scenes, you get the magic box: any subject, in any scene, anywhere, at photorealistic output no human can match for time, convenience and money.

However, accountability and repeatability are some of the components that make up “taste”. This, and judgement, is being able to discern, knowing that the weight of a decision carries consequence, which is why a human navigating these decisions during production is pivotal for brands that know their own story and want to explore AI without it becoming slop.