Generating at full pixel resolution would be ruinously expensive, so most systems compress first. An encoder turns the image into a compact numeric form, all the generation work happens there, and a decoder expands the finished result back into pixels. The heavy lifting happens on something a fraction of the size of what you see.
The compression is lossy and its fingerprints are visible if you know where to look. Very fine texture, small text and single-pixel detail live below what the representation can hold, which is part of why tiny lettering is unreliable across an entire family of models regardless of prompt.
It also explains why an output at a strange size can look subtly wrong. The encoder works in fixed blocks, so dimensions that do not divide cleanly get handled awkwardly at the edges — which is why models publish specific supported sizes rather than accepting any number you type.
In practice
- Generation happens compressed; the picture you see is decoded at the end.
- Detail finer than the compression can carry will not survive, whatever the prompt says.
- Supported output sizes exist because the encoder works on fixed blocks.
The mistake to avoid
Blaming the generator for soft micro-detail. Some of that loss happens in the encode/decode round trip, before generation has any say.
Related terms
Diffusion model
A diffusion model generates by starting from random noise and removing a little of it at a time until a picture or clip is left behind.
Resolution
Resolution is how many pixels an output contains, usually named by its height — 720p, 1080p, 4K — and set before generation rather than after.
Sampling steps
Sampling steps is how many passes a model takes to turn its starting noise into a finished output — more passes, more refinement, more time.
Upscaling
Upscaling raises the resolution of media you already have, inventing plausible detail rather than recovering detail that was never recorded.
Diffusion transformer
A diffusion transformer is a diffusion model whose internals are a transformer — the same architecture behind large language models — instead of the convolutional network earlier image models used.
The all-in-one AI studio for creators. 60+ models for video, image, voice, music and lipsync in a single app.