You upscale a clip, watch it full screen, and it looks great. Then you zoom to 100% to check a face, and it looks worse than the original. Waxy skin, painted edges, patterns where there was noise.
This is not a bug in the tool. It is the difference between resolution and detail, and it is measurable.
A neural upscaler does not invent detail. It predicts, for every new pixel, the most probable value given its neighborhood.
At 2x, one pixel becomes four. Three are predicted. The ratio of measured to predicted information is still workable.
At 4x, one pixel becomes sixteen. Fifteen are predicted. Beyond that point the model stops looking at your image and starts reproducing patterns it memorized during training.
That is where the artifacts come from: uniform texture where there was grain, glassy skin, edges that look drawn rather than photographed.
Preset NATURAL HD. Left is the blurry original, right is the result. Skin texture is preserved.
Your eye cannot separate resolution from detail. A larger image always reads as sharper at a glance, because there is simply more signal reaching the retina.
At 100%, that advantage disappears. You see one screen pixel for one image pixel, and the prediction errors become visible.
So the only valid test is: same crop, same zoom, side by side, at 100%. Anything else is measuring file size, not quality.
Most people get the order wrong, then blame the model.
Step 1: evaluate the damage. Open at 100%. If the noise is compression (8x8 blocks), you need cleanup first. If it is film grain, aggressive cleanup will erase the original texture along with it.
Step 2: denoise, gently. Always less than feels necessary. Whatever noise remains is part of the document; the noise the algorithm adds is not.
Preset SOCIAL DENOISE. Left is the re-downloaded clip, right after cleanup. Block artifacts are gone.
Step 3: upscale 2x, not 4x. From 720p to 1440p. Going to 4K from 720p means you are asking the model to invent 93% of the pixels.
Step 4: sharpen last, and lightly. Sharpening before upscaling means the algorithm amplifies the halos you just created.
Preset CLEAR REALITY. Structural detail survives; only compression noise is removed.
Not the face alone. Check these:
If hair looks right but fabric turned into a flat patch, the mode was wrong, not the strength. Preset TRUE TEXTURE. The repetition should stay regular, with no stretched or compressed areas.
I use UpgoLAB in the browser. Modes are separated by material type, and the denoise, upscale, sharpen order is already wired internally — which is the exact part most people get wrong by hand. The mode list shows which one is meant for which material. Free to try, no install.
Being straight about the limits:
These tools extract the maximum from what exists. They do not create new material.
The order of steps changes the result more than any single setting.