Upscale and retouch AI images without the plastic look
Why upscaled output reads as fake, and the finishing pass that fixes it.
Before you start
- A generated image at native resolution
- Any editor with layers and opacity
What you will be able to do
- Choose an upscaling factor that does not invent texture
- Restore grain and micro-contrast after enhancement
- Recognise the three artefacts that give an upscale away
A generated image often looks fine at preview size and wrong at full size — skin like plastic, fabric with no weave, an overall glassiness.
That is usually not the generator. It is the upscaling and enhancement stage inventing detail that has no texture.
Upscale less than you think you need
Every doubling is more invented detail. Two is usually plenty.
An upscaler does not recover detail — it fabricates plausible detail. At 2× that fabrication is mostly convincing. At 4× and beyond it starts producing the characteristic smooth, slightly wrong surfaces, because there is far more inventing to do.
Work out the size you actually need first. Most people upscale to a number that sounds impressive rather than one anything requires.
Blend the upscale back against the original
Full-strength enhancement is where the plastic comes from.
Put the upscaled version over a plain resize of the original and pull the opacity back to somewhere around 60–80%. You keep most of the sharpening and recover the original's irregularity.
Mask it where it matters: faces and fabric usually want less enhancement, hard edges and architecture can take the full effect.
Put grain back
Real images have noise. Perfectly clean surfaces read as synthetic.
A light monochrome grain over the whole image does more for believability than any amount of extra sharpening. Keep it subtle enough to be invisible at a glance and present enough to break up the flat areas.
Uniform noise across the frame is the goal — grain that follows the subject looks like an effect rather than a photograph.
- Add grain last, after any resizing. Scaling an image after graining it smears the grain into something that reads as compression.
Check the three usual artefacts at 100%
Repeating micro-patterns, melted small text, and over-defined eyes.
Zoom to full size and look specifically for: repeating texture in areas that should be irregular, like foliage or gravel; small text and logos turned into confident nonsense; and eyes, which upscalers over-enhance until they gleam unnaturally.
All three are fixable by masking the upscale back in those regions rather than by re-running the whole image.
Upscale conservatively, put grain back, and let some things stay soft. The eye reads uniform sharpness as fake.
Common questions
Was this guide useful?
87% of readers found this useful
Read next
Get consistent characters and style in AI image generation
Generating one good image is easy. Generating twelve that clearly show the same person in the same style is the actual problem — a…
Write image prompts that get the composition you meant
Image models are good at subjects and bad at guessing layout. Saying the composition out loud is what closes the gap between what…
Write SQL with AI against a schema it has never seen
A model will write syntactically perfect SQL against a schema it is guessing at. The query runs, returns numbers, and the numbers…
Use an AI coding assistant without shipping code you did not read
AI assistants are genuinely good at writing code and genuinely bad at knowing when they are wrong. This is the review process that…
