How we built Premium AI Upscale

1. Same image, same regions. We pick the parts of an image where upscalers disagree the most: the face, the eyes, the lips, large areas of skin, hair against the background, text and logos, and small textures like water droplets, sequins or jewellery.
2. Zoom to 100%. At normal size most upscalers look alike. The differences (cracked skin, invented lashes, tile patterns) only show up close.
3. Keep the original next to every result. It is the only fair reference. On one test cover, a glow around the model's face looked like an upscaling artifact; the original showed it was part of the design.
4. Measure, then look again. Every claim in our tables is backed by a number, and every number is checked by eye.
5. Rank by fixed rules, in this order:
   1. Nothing invented: no texture, lashes, glitter or patterns that are not in the original.
   2. Small features stay true: eye reflections, lip colour and skin tone match the original.
   3. Sharpness, only after the first two. A sharper result that invents detail ranks lower.
   4. Consistent across the whole frame: no seams, tile patterns or patches.
   5. Cost: time and memory.

The measurements, in plain words

• Eye reflection. How much brighter the small reflection in the eye is than the iris around it, on a 0-255 brightness scale, measured where the original has it. Close to the original is right. Far above it means the reflection was overblown.
• Fine-detail score. The average difference between an image and a slightly blurred copy of itself. Higher means more fine texture. Invented texture also scores high, so this number is never read on its own.
• Glow check. Whether the original has a see-through fringe around the person (typical of a pasted cut-out). We measure how much of the person's colour leaks into the background at each pixel of distance from the edge. A designer glow reads about 0.3 at 2 px; a natural photo edge reads about 0.02.

Some uploads are larger than the detail they hold. A frame that was enlarged before upload, or a phone photo whose noise reduction smoothed the texture away, looks soft at any size, and a 4x on its own only enlarges that softness. For these images the pipeline adds one step before the 4x by default: the detail pass. The whole-image restorer runs at 1x, at the upload's own size, and rebuilds edges and texture the blur hides. The 4x then enlarges crisp content instead of blur.

Which images get it. We shrink the upload to a quarter of its size, enlarge it back, and compare the detailed areas with the original. If almost nothing changes (above about 45 dB), the upload holds much less detail than its size suggests. Until the threshold is proven on many more images, a person confirms each case.

The detail pass is the same step in both tiers. Fast runs its own 4x after it and Premium runs its own; neither tier's result is ever used for the other.

Where to see it. Comparisons 7 and 8 above are the two test images that get the detail pass: their Premium and Fast rows are the final results with it, and their last two rows show what each tier would have delivered without it.

We list these openly, because a tool that hides its weak spots cannot be trusted on the rest:

• Out-of-focus areas stay soft. Premium follows the photo's depth of field, so blurred tiles or bricks stay blurred, where Fast sharpens them (outdoor photo).
• Smooth surfaces can gain invented texture: fibres on a printed towel, flecks in smeared paint, streaks around a skin highlight, and lace extended across a panel that only hints at it (outdoor photo, dim phone photo, upside-down face).
• Video compression noise turns into a pattern. Faint colour blotches on a wall become crisp pink dashes (video still).
• Already-enlarged or heavily smoothed images gain little from a 4x alone. The detail pass fixes most of it, but adds detail of its own: flyaway hair strands, and redrawn letters where text was too soft to read.

1. Colour. The upload is converted to sRGB, honouring any embedded colour profile.
2. Whole-image restoration. A 3-billion-parameter one-step diffusion restorer produces the full 4x image in a single step. The image is encoded into a latent space 8x smaller per side; the encoder and decoder work in 512 px tiles.
3. Face detection. Each face is located with its outline, eyes, pupils and lips.
4. Face restoration. A 1.4-billion-parameter one-step restorer processes the frame at 4x in 512 px tiles with 64 px overlap, and the face region is cut from its result.
5. Blend. The outline mask, extended 8% of the face height above the brows, with a 20% fade inside the outline and a colour match in the fade ring.
6. Glow check. A segmentation mask of the person, and the fringe measured 1-12 px outside the edge. Removal runs only above the thresholds (0.15 at 2 px and 0.05 at 4 px), by un-mixing the person's colour from the background.
7. Output. Exactly 4x the original width and height.

Most orders hold ten images or more, so what matters is how long the whole order takes, not one image.

Every image in a Premium order goes through two large restoration models, and photos with faces get the face pass on top, so a typical order takes hours. Larger images and larger orders take longer; an order of a few small images is ready sooner. We are working on making Premium faster.