Image Generation for Concepts

Concept design is the most divergent stage of a project: the goal is to open the solution space, not close it. You want many rough directions to react to, cheaply and fast. That is precisely the shape of task diffusion models are built for. Stable Diffusion, Midjourney and Adobe Firefly all take a text prompt and produce fresh images by iteratively denoising random noise toward something that matches your words - which means every run is a new roll of the dice, and variety comes for free.

The honest framing is that image AI is a fluency amplifier for the early hours of a project. Where you might once have sketched five thumbnails in an afternoon, you can now look at fifty in twenty minutes - enough to notice a move you would not have drawn yourself, or to feel quickly that a whole direction is wrong. That is real value, because reacting to options is often easier than generating them from a blank page.

The flip side matters just as much: this is a low-stakes, high-volume use, so your scrutiny can be loose here - you are hunting for a spark, not a buildable answer. As the project converges, image AI's role shrinks. It cannot resolve a plan, honour a dimension, or keep a column grid straight. Treat its concept images as evocative sketches, never as drawings, and you will use it for what it is genuinely good at without being misled by how finished it looks.

There is also a plain economics to this. A generation costs cents and seconds, where a hand-built moodboard or a physical study model costs hours. That collapse in the cost of a single option is what actually changes the workflow: when an option is nearly free, you can afford to be wrong forty-nine times to be surprised once, and you stop over-investing in your first idea. But cheap options carry a subtle danger - the output looks resolved, glossy and confident, which tempts you to treat a lucky picture as a decision. The professional habit is to separate the two feelings: excitement at an idea, which is welcome, from trust in an image, which must be earned. Early on, feel free to be excited and slow to trust.

A weak prompt gets weak, generic output; a framed prompt gets useful output. The difference is not magic words - it is saying what you actually mean. A reliable structure for architectural and interior concepts fills six slots: subject, material and form, light and mood, view and camera, medium and style, and constraints.

Compare the two:

Weak:   modern house
Better: small hillside house in rammed earth and timber, low
        overhanging roofs, soft morning light, calm mood,
        eye-level wide-angle view, architectural photograph,
        no people, no text, no signage

The second is not longer for its own sake - each clause is a decision you are making instead of leaving to the model's averages. Naming a medium ("architectural photograph", "physical massing model", "loose watercolour concept") is one of the highest-leverage moves, because it changes the entire character of the output. So does the camera - "eye-level", "aerial", "one-point interior" - which quietly controls how architectural the result reads.

Each tool has its own dialect. Midjourney rewards evocative, comma-separated phrases and offers parameters like --ar 3:2 for aspect ratio and --stylize for how much it embellishes. Stable Diffusion (via ComfyUI or Automatic1111) exposes seeds, samplers and a negative prompt where you list what to exclude. Firefly is trained on licensed and public-domain content, which makes it a safer default for commercial concept work. Learn one tool's dialect well before spreading yourself thin - the six-slot thinking transfers across all of them.

Two more habits sharpen the framing. First, prefer concrete nouns and qualities over vague praise: 'deep window reveals in board-marked concrete' directs the model, while 'beautiful stunning masterpiece' merely nudges its style dial and tells it nothing about your building. Second, feed it references where the tool allows - Midjourney and Stable Diffusion both accept an image as part of the prompt - so the direction starts from something you chose rather than from the model's defaults. This is also where a good LLM earns its place beside the image tool: describe your project to ChatGPT, Claude or Gemini and ask it to draft five richly specified image prompts, then edit them. That chaining - words tool feeds image tool - is a small taste of the pipelines Module 8 builds, and it consistently produces better-framed prompts than typing blind.

The beginner's mistake is to rewrite the whole prompt every run, so you never learn what any change actually did. The disciplined habit is to vary one slot at a time and keep the rest fixed - change only the material, or only the light, or only the camera - and watch what moves. That turns generation into a controlled experiment rather than a slot machine.

Most tools give you levers to iterate tightly once you have something promising. Midjourney's variations nudge a chosen image; its seed lets you reproduce a look and change one word against a stable base. Stable Diffusion's seed does the same, and lowering the denoise strength on an image keeps its bones while restyling the surface. Learning to hold a seed steady is the difference between exploring around an idea and starting over every time.

A simple, repeatable iteration loop:

1. Write a six-slot prompt; generate a batch of ~8.
2. Pick the 1-2 that have a real idea (not the prettiest).
3. Lock that seed. Change ONE slot. Generate again.
4. Repeat until the direction is clear - then stop.

Notice the stopping rule. Divergence is addictive; it is easy to generate for an hour and end with a pretty folder and no decision. Set yourself a budget - a number of batches or a timer - and when it is up, converge.

It helps to iterate with intent rather than hope. Before a run, say out loud what you are testing - 'does a heavier base ground this form?' - so each batch answers a question instead of just producing more pictures. Keep a light record, too: save the seeds and prompts of your keepers, because a look you loved and cannot reproduce is worse than never having seen it. A tidy folder of 'prompt, seed, one-line idea' turns an afternoon of generation into a small, reusable library you can return to when the project evolves - and it makes the eventual hand-off to your own modelling tools far cleaner.

Generation is the easy half; curation is the design work. The model has no idea which of its outputs serves your brief, your site, or your client - only you do. So the real skill on display in a good AI concept phase is not prompting; it is ruthless selection and interpretation. From fifty images you might keep two, and even those you keep for a move - a sectional idea, a material pairing, a quality of light - not as finished proposals.

Judge on the idea, not the polish. Diffusion output is seductively rendered, and it is easy to fall for a beautiful image that says nothing. Ask of each keeper: what is the one architectural or spatial idea here, and is it any good for this project? Then extract that idea and take it back into your own tools - trace over it, model it, sketch from it. The image was a prompt for your thinking, not a deliverable.

A quiet risk to name: image AI can smuggle in someone else's language. Prompting a named living architect's style, or leaning on the model's defaults until everything looks the same, produces derivative work that is neither yours nor defensible. Use the images to provoke your idea, then redraw it in your own hand. That last step - reinterpreting rather than copying - is what keeps the concept authored by you, which is the whole point of this course.

A useful test for whether you have curated well: can you defend each keeper in a sentence to a sceptical studio-mate? 'I kept this one for the way the roof peels up to bring north light deep into the plan' is a design reason; 'I kept it because it looks cool' is not, and it will not survive a review. Naming the move also forces you to notice when a whole batch is empty - technically competent but saying nothing - which is your signal to change the prompt, not to keep scrolling. Curation, done honestly, is where a pile of AI images quietly becomes the beginning of a real project.