CLARITY FIELD NOTES
Understanding 8× AI image upscaling
Eight times the width and height means 64 times the pixels. An 8× result is a demanding operation, even though the model itself is small.
What Clarity actually does
Both bundled models are native 4× networks. For requested output above 4×, Clarity runs 4× AI on overlapping source tiles, downsamples that result to 2×, then runs the 4× network again. This produces an 8× AI pipeline followed by a final resize to the chosen dimensions. A 6× request also uses two passes and finishes at 6×.
A concrete memory example
A 1000 × 1000 source becomes 8000 × 8000 at 8×: 64 million pixels. Just one RGBA canvas needs about 256 MB in decimal units. The source, tile tensors, working canvas, comparison, encoder and GPU allocations require additional memory. Clarity’s 40 MP output guard intentionally rejects this example; low-memory mode uses a 16 MP output guard.
Tiles limit inference memory
Rather than passing the entire source through the network at once, the worker builds overlapping tiles, infers each tile and discards its outer halo before stitching the center. This reduces boundary effects and peak inference size. It does not eliminate the memory needed for the final image or guarantee a seamless result for every source.
Choose the smallest useful scale
If your print needs 3508 pixels on the short edge and your source already has 1800, a 2× enlargement may be enough. Extra passes add time and can amplify invented detail. Inspect the comparison at 100% and avoid assuming a larger file means a better result.
If the run cannot finish
Crop to the area you need, lower the dimensions, try low-memory mode or switch engines. Cancel terminates the whole worker rather than waiting for the current network evaluation. Extremely large exports remain limited by canvas support, encoding memory and available device resources.
Open the free image studio · Calculate print size