Guide
Why your AI image looks soft, and whether upscaling fixes it
Softness and small size are different complaints with different cures, and reaching for the resolution slider fixes only one of them. Often it is not the one you have.
The distinction
Resolution is a count. Sharpness is a contrast.
Resolution is how many pixels there are. Sharpness is how abruptly tone changes across an edge. A four-megapixel image with crisp edges looks sharp; a sixteen-megapixel one where every edge fades over several pixels looks soft, and enlarging it further only spreads the fade across more of them.
Generated images tend toward the second. A model producing a plausible average has no reason to commit hard to any particular edge — the average of many possible edges is a gradual one — so detail is distributed evenly and nothing is decisively sharp.
That is why asking for a bigger image so often disappoints. The complaint was never about pixel count.
The actual fix
Give it a reason to put detail somewhere
Real lenses cannot focus on everything. There is a plane that is sharp and everything nearer or further falls away, and the contrast between those two things is a large part of what the eye reads as sharpness. An image sharp everywhere has no such contrast, so it reads as sharp nowhere.
Naming a focal length and an aperture gives the model that structure. 85mm, f/1.8, focus on the eyes tells it to commit detail to one plane and let the rest go, which produces an image that looks considerably sharper than a uniformly rendered one at the same pixel count.
Texture helps for the same reason. Skin pores, fabric weave, grain in wood — concrete surfaces give edges to resolve. So does lighting with direction, since a flat-lit subject has little tonal variation for an edge to be made of; that argument is in the lighting guide.
Upscaling
What it adds, and what it cannot
An upscaler does not recover detail, because there is none to recover. It invents detail consistent with what is already there — which is genuinely useful, and worth being precise about.
It works on an image whose structure is already right: correct anatomy, sound composition, the light doing what it should. Invented pores on a good face look like pores.
It does not work on an image with a structural problem. Upscaling a hand with six fingers produces a larger six-fingered hand rendered more convincingly, and the extra conviction makes it worse rather than better. Fix the generation, then enlarge.
It is also a separate pass with a separate charge on most platforms, which is worth knowing before building it into a routine — see what generation actually costs.
A trap
Generating as large as possible usually costs more and looks worse
Two reasons, and both argue the same way.
Pricing follows pixels. Image generation is commonly billed per megapixel, so a 4K request is not a little more expensive than a 1K one — it is many times more, for every attempt including the ones you discard.
And models drift outside the size they were trained at. Pushed well beyond it, they start repeating what they know: a second set of limbs, a duplicated face in the background, a pattern that restarts halfway down. These are not rendering faults but composition faults, and they appear exactly when the canvas is larger than anything the model saw.
Generating at the native size, choosing the frame that works, then upscaling that one is almost always cheaper and usually better. And since each run lands somewhere different, the cheaper the attempt the more attempts you can afford.
Asked often
Short answers
- Why is my AI image blurry or soft?
- Usually it is not resolution. A generated image can be large and still soft, because the model spread its detail evenly instead of putting it where a lens would - there is no plane of focus, so nothing is truly sharp. Asking for a focal length and an aperture gives it a reason to make one part crisp and let the rest fall away, which reads as sharpness even at the same pixel count.
- Does upscaling make an AI image better?
- It makes it bigger, and it invents the detail it adds. That is genuinely useful on an image whose structure is already right, and useless on one whose problem is composition or anatomy - upscaling a wrong hand produces a larger wrong hand in more convincing detail. Fix the generation first and enlarge second.
- Should I just generate at the highest resolution available?
- Not automatically. Generating larger costs proportionally more because pricing follows pixels, and many models were trained at a particular size and drift at much larger ones - repeated limbs and duplicated features are a common sign of pushing past what a model handles. Generating at the native size and upscaling afterwards is often both cheaper and better.
Keep reading
The rest of the series
- Using your own photo as a referenceWhat a reference carries, how many to supply, and why one came back unchanged.
- Why a prompt gets refusedThree different refusals wearing the same error. Which one you have, and which can be reworded.
- Does the language of your prompt matter?The script matters more than the language, and the gap is larger than anyone tells you.
Where this comes from
Unlimited Studio is built around this one problem
Everything above is what the product does: a character is locked from eight reference angles, each generation conditions on whichever angle its prompt implies, a score checks the result against the reference, and a film is shot one frame-chained shot at a time. What it will and will not generate is set out in the content policy.