3_Writings_2_90% rule V2

There's a point where generating again stops helping. Knowing where it is, and finishing the rest by hand, is the line between using AI to lift a brand and a brand becoming synonymous with slop.

When is a generation good enough that running it again won't get you any closer to what you pictured?

I call it the 90% rule. It's the point where I lock a generation and stop asking the model for more. Where that point sits comes from years in image production, retouching and pipelines, and knowing how long a fix will take by hand, measured against what I've learnt about how these models behave.

That's what taste and judgement are: a skill set built from human experience. A model doesn't have it. It's working to satisfy goals shaped by its training, and it's good at that. The result is often convincing, and often empty. It satisfies an algorithm. It doesn't satisfy people who expect something real from the brands they love and trust.

The gap is widest in complex scenes, where geometry sits behind more geometry. An interior with walls, furniture and depth, not a still life with one plinth and one object on top.

How I know I'm at 90%

I iterate until I've circled the idea as closely as I can. Then I run the prompt once more across the models that performed best along the way. Sometimes four or five times, looking for the unicorn: the 100% generation.

It rarely turns up. In everything I've learnt so far, there's always clean-up of some kind. That clean-up is the line between slop and something curated by a person, for people.

Once a prompt is at 90%, the model is sampling from very similar territory, so the results come back very similar too. These are 90% generations from my OK. project.

With the images below, I wanted dynamic lighting and poses, varied subjects, and a scene far enough from the symmetry that gives slop away. The popular wide lens, low angle was used as a dynamic factor. Running the same generation with the same inputs and prompt usually returns much the same image. So knowing when to rerun against a checklist, and when to change the prompt instead, is one of the most valuable skills I've picked up.

3D · Blender
AI Image Model · ChatGPT Image 2 · Medium

Above, the 90% generation. Below, the final after retouching.

What I'm looking for

At 90%, the questions change. What will be easy or hard to inpaint? What can I fix myself with traditional retouching? Some models are better at inpainting certain things than others, and it's usually the hero object or subject that decides whether a 90% generation is worth keeping.

3D · Blender
AI Image Model · MODEL NAME

Why not just keep inpainting over the same image?

Inpainting, however small or isolated, adds destructive artefacts every time the image goes back through. Fixing one hand in a group shot still means the model redraws that area from scratch. That's the nature of a diffusion model: the process is probabilistic, so every pass is a new roll. Text degrades most along with repetitive lines or anything with texture.

Some of that drift in light, colour and composition can be reduced in the prompt itself. I put the negatives at the bottom. With reasoning models, that lets the model work through the creative, positive sections first, without being interrupted by negatives at the start or confused by negatives mixed in among them. More on that in how reasoning models read a prompt.

PLACEHOLDER — the same area across inpainting passes, and what drifted.

Then there's cost

Cost is worth weighing whenever a final image needs several single passes. It's easy to get lost in the pile of generations: downloading, dropping pieces into the hero image, painting them in or overlaying them. I go into this further in what AI image models actually cost.

Getting from 0 to 90%

The rule only works if you reach 90% in the first place. This is the setup that gets me there.

Work with a chatbot. Use the smartest one on your paid plan, and use features like Claude's Projects: self-contained spaces with everything the chat can do, but without memories from your other chats about cats, baking or deadlifts leaking in. Set up the brief and the rules there and let it guide the prompts.

Test more than one model. Write a prompt diverse and demanding enough that you can run it through five models and choose between the results. A node workflow works well here. Upload one reference image or text prompt, connect it to several models and run them all at once. It feels like orchestrating, and in some ways it is.

Test characters before you commit. If there's a character story, run test character sheets that fit your decisions first. How I approached this for Mirage is in likeness checks.

Don't ask it to judge the image. A chatbot is great at refining prompts. It isn't good at looking at images. Describing the look and feel of an image or video is still a human job. More on that in what models see and what we see.