The 90% rule, and 100% without re-rolling

When is a generation good enough that running it again gets you side steps, not steps forward to 100%?

I call it the 90% rule. It's the point where I lock a generation and stop re-rolling the image or motion prompt. Some of the basis for this comes from my years in image production, retouching and pipelines. Knowing how long a fix will take by hand, measured against what I've learnt about these models' strengths and weaknesses and how they're trained, helps me decide whether to move into the next phase.

The problem amplifies as scenes rely more on the text prompt to drive the generation. The other way would be inpainting, which is far easier for a model to achieve, and offers a better, higher-quality asset than re-rolling. An example would be a subject with eight fingers on their left hand. There's more prep in cropping the hand out and saving it to your desktop, which already has 100 screenshots, uploading it, writing an editing prompt. But the result has a higher chance of working if it's re-uploaded with only the hand as input.

Focused inpainting with separate prompts or separate sections of the image is time-consuming, and defeats the purpose of using magical AI, but I've found that issues amplify the larger the inpaint task is. The gap is widest in complex scenes, where geometry often sits behind more geometry. As an example, an interior with walls, furniture, windows, decor objects and depth introduces more variables than a still life with one plinth and one object on top. Variables are opportunities for drift, and too much drift is AI slop.

I’m philosopher for moment: this 90% point is part of what defines the difference between a human and AI. It is a component of taste and judgement: only a human can define it, through a lived experience of how a decision has consequences. A model mostly has signal from humans rating images in training (RLHF) and a black box no one can fully explain (one for another post). 90% > 100%, I believe, is a skill set built, so far, only from human experience.

We all want the magic box without the slop.

How I know I'm at 90%

It's usually around a dozen... or two dozen generations in. I iterate until the prompt's structure works, and adjustments to parts of the prompt start delivering sideways results that look mostly the same on each re-roll. Then I sometimes run the prompt a few more times across the models that performed best along the way, looking for the zero-inpaint generation: the 100% generation. The unicorn.

It rarely happens. In everything I've learnt so far, there's always clean-up of some kind. I call it the anti-slop human touch. That clean-up is the line between slop and something curated by a human. They used to call it “the eye” in photography; I just see it as caring too much about the details – 500% zoomable.

I’m dork for moment: once a prompt is at 90%, the model is sampling from very similar territory within what's called its “latent space”, so the results come back very similar. That's why it goes sideways. The model is sampling from the same area, and that mathematical space is code, but it isn't deterministic. Dragging a paintbrush across the canvas with the same settings in Photoshop always gives the same result. A model is probabilistic. I find this fundamental to understanding the core difference between a human with a program and the magic box of AI.

Pictures are so much better than words about science

Below are some 90% generations from my OK. project.

At this stage, all the product images for the different flavours had been generated and locked, and were the only references supplied with the image prompt. With these generations, 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 look was the art direction goal. The more popular the visual direction of an image or motion clip, the more training data there is, and the more commentary on that style to train AI chatbots to help people describe it in prompts. So something like this, which has been done quite often, with the core idea of subjects holding an object, is high signal for AI models.

What I'm looking for

At 90%, the questions are simpler for how to proceed. 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 informs whether a 90% generation is worth keeping, because a background object is both easier to replace and less likely to be scrutinised by a viewer than the hero object. By background I don't mean a window or wall shifting from image to image – which is common.

Below is an example of how balancing the key storytelling components of an image is the most important milestone I aim to cross, when generating endless iterations is so easy. Once the prompt included all the major components I was looking for, I ran the same prompt five more times, looking for: A) overall scene composition – what the mountain looked like, whether the town sat in a realistic enough location in the scene, and whether the bottles were scaled correctly; and B) whether the two little kids had the right facial expressions and body language to convey the wonder of seeing such mythical, colourful water bottles at Christmas.

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 can still mean the model redraws the entire image. That's a diffusion model: the process is probabilistic, so every pass is a new roll. Text degrades most, along with repetitive lines and anything with texture. Exposure and contrast shifts are also common depending on the model used.

Newer models are becoming much better at masked edits without introducing new issues. However, I've found every model introduces some kind of change as inpainting continues, iteration over iteration. Common issues I find are the introduction of noise or banding, objects moving or rotating, or scene elements changing completely.

This is why I measure the workflow I use based on the subject, its use, how much training data I think there is on that subject, and whether the subject itself inherently has variance. An example would be humans: there are billions of images of every kind of human subject in training, and humans are inherently not symmetrical, so variances and artefacts are almost something you'd want in a generation. When compared to an object with a very structured, solid, set geometry like a water bottle, audiences expect that bottle not to change in mass or scale once they see the first image of it.

Some of this drift in light, colour and composition can be reduced in the prompt itself using structured negatives towards the bottom of the prompt. With reasoning models, that lets the model work through the creative, positive sections first, without contradictory negatives mixed into positive areas.

Below is an example of a 90% starting point that was then used for a seven-pass inpaint pipeline, where each output image was added as the input file for the next inpaint pass. By pass 3 there was exposure drift, by 5 banding, and by 7 the image is 100%, but at the cost of new issues which are in some ways harder to fix.

Then there's cost

Cost is worth weighing whenever a final image needs several single passes. Depending on final use, a pipeline can require dozens of generations: downloading, dropping pieces into the hero image, painting them in or overlaying them. This adds up in both time and credits used. I go into this further in Nothing is free – navigating model cost.

100% in a time when output is endless

Endless output that needs to move through creative teams and wider touchpoints to its final use environment requires some kind of iterative pipeline that can be easily scrutinised and adjusted.

The perception is AI is magic, and it is, until it's not. Then it becomes quickly apparent that the time that was thought to have been saved by running production through AI becomes a never-ending inpainting task, which can feel endless when a product needs to look the same in every image. These are some of the challenges I’ve had to learn to navigate in creating my AI projects, and relevant to businesses which are aware of the output, time and budget savings, but maybe not so when it's at the expense of your product's accuracy.

This 90% article is part of a larger piece I'm writing, “What is AI slop”, where I explore this from a creative perspective. That's where I started, in traditional creative, and I've moved with AI since February 2025.