T/07 · UPSCALE AI

Upscale images. Synthesize real detail.

Up to 4× with a model trained on billions of real image patches. It doesn't stretch pixels — it reconstructs plausible texture, edges and gradients. Old photos, thumbnails, screenshots — print-ready.

Drop a photo to upscale

or drag and drop here

JPG · PNG · WebP · AVIF · up to 50 MB

S/02 — HOW IT WORKS

Three steps. No detours.

Upload, process, download. Files are auto-deleted 24 hours after processing.

S/02 · 01
compress · selected

Choose a tool

Pick what you need from the toolkit.

S/02 · 02

Drop your image

Drag, click, or paste. No upload bar required.

S/02 · 03

Download

Result is ready, original untouched.

S/SPECIMEN — SYNTHESIZED DETAIL

From 480 px to 1920 px. No paste, no pixelation.

Pixel-level comparison between standard bicubic interpolation and our super-resolution model. The output isn't a stretch of the source — it's detail reconstructed from learned patterns.

⊕ S/SPEC · LOWRES.JPG · ID 0E9D
480²

INPUT · 86 KB

AI · 4×

1920²

OUTPUT · synthesized

480 × 320 → 1920 × 1280 · 4× linear · 153 600 px → 2 457 600 px (16× px) · imgpix-sr-v1

S/V — WHY UPSCALE

Upscale, four angles.

T/07.1 · SCALE FACTOR MATRIXFROM 480 × 320

SINGLE PASS

480 × 320960 × 640

4× px

DEFAULT

480 × 3201920 × 1280

16× px

multi-pass

CHAINED

480 × 3203840 × 2560

64× px

multi-pass

CHAINED

16×

480 × 3207680 × 5120

256× px


4× lineal = 16× píxeles · 1 sola pasada

SCALE

Up to 4× in a single pass.

A 512² photo becomes a 2048² print-ready file; a 1080² thumbnail turns into a 4320² hero. Optional multi-pass to push further without sacrificing quality.

T/07.2 · PIXEL DIFF · 12 × 12SAME REGION · SAME OUTPUT SIZE
BICUBICinterpolated

EDGE: smooth

TEXTURE: averaged

DETAIL: lost

AI · SR-V1synthesized

EDGE: crisp

TEXTURE: reconstructed

DETAIL: synthesized


Misma región · misma resolución de salida · diferencia a nivel píxel

DETAIL

Synthesized detail, not interpolated.

Bicubic stretches gradients and blurs edges. Our model reconstructs plausible texture, corners and micro-detail from learned patterns. The difference is visible at pixel level.

T/07.3 · USE CASESRESCUE PROFILES
FAMILY SCAN

Escaneos vintage

600 × 400 · grano

2400 × 1600 · texturas recuperadas

WEB THUMB

Miniaturas web

256 × 256 · fuente única

1024 × 1024 · listo para hero

STOCK PRINT

Stock para print

800 × 600 · web stock

3200 × 2400 · ready-to-print

JPG ARTIFACT

JPG comprimido

~70% quality · ringing

limpio · sin ringing


Donde un upscaler clásico falla · estructura recuperada

RESCUE

Rescue impossible photos.

Grainy family scans, forgotten web thumbnails, low-res screenshots, over-compressed JPGs. Where a classic upscaler fails, the model finds structure.

T/07.4 · MODEL SPEC SHEETimgpix-sr-v1

imgpix-sr-v1

super-resolution · frozen weights · v1.2.0

LIVE
MODEL
imgpix-sr-v1
super-resolución · frozen · versionado
TRAINING
~3 B parches
fotos reales · sin tus uploads
INFERENCE
8–25 s típico
GPU dedicada · escalado a carga
MAX INPUT
4 096 px lado largo
JPG · PNG · WebP · AVIF · HEIC · TIFF
PRIVACY
Zero retention
tus imágenes no entran al pipeline

MODEL

Engine spec sheet.

imgpix-sr-v1 model, trained on billions of real image patches. Inference on dedicated GPU, typical latency 8–25 s. Your images never enter the training pipeline.

SPECT/07
SCALE
2× · 4× (default) · up to 8× with multi-pass
MODEL
imgpix-sr-v1 · super-resolution
INPUT
JPG · PNG · WebP · AVIF · HEIC · TIFF
MAX INPUT
Up to 4096 px on the long edge
LATENCY
8–25 s typical, depending on size and load
PRIVACY
No training on your images

S/F — FAQ

Upscale, briefly answered.

How much can I upscale?
Up to 4× linear in a single pass (16× pixels). Beyond that, chain 2× passes: 2× → 4× → 8×. Each pass adds inference time and cost; the model is optimized for the 2×–4× range.
Does it add real detail or just enlarge?
It synthesizes plausible detail from local context. It looks real, but it's learned reconstruction — not hidden information from the original. For forensic use, frame it that way.
Best for which images?
Natural textures (skin, fabric, foliage, landscapes) is where it shines. Synthetic graphics and screenshots also work well. Heavily compressed JPGs can amplify artifacts — a denoise pass first helps.
Are my images used to train the model?
No. Your uploads are used exclusively to process your job. Zero feedback into the training pipeline. The model is frozen and versioned.