A blurred, low-resolution photo enlarged four times by both upscalers side by side: the faithful one and the creative one. The full workflow, no account needed.
Real workflow, real results. Pan and zoom freely.
Use this workflowThe real workflow: the same 360-pixel family photo feeds both upscalers, and the canvas shows what separates them.
At a glance
An old phone photo, a botched scan, an image pulled out of a chat thread: three hundred pixels wide, motion blur, compression blocks. This workflow enlarges it four times. The photo enters the Media node and feeds two Enhance nodes, each set to a different upscaler.
The two do not do the same job, and that is the whole point of the template. Topaz Precision stays faithful: it recovers the detail that exists in the source, keeps the faces recognisable, and leaves artefacts wherever the source was smooth, jumpers and fabrics. Topaz Bloom 2 reinvents the detail: the photo is superb, materials and background rebuilt, but the features drift.
On the demo photo, measured on 1 September 2026: Bloom 2 changes the father's jaw, adds an earring to the mother and turns the boy's jumper into clean stripes. Splendid, and no longer quite the same family. For a family photo, likeness comes before beauty; for a background or an object, the trade-off flips. Both enlargements cost forty-two cents.
Import the damaged photo
Low resolution, blurred, compressed: that is the use case, not a problem. The Media node shows the source dimensions, and the Enhance node announces the output size before generating.
Run the faithful upscale
Topaz Precision at 4x. This is the one to try first, and often the only one you need: it works from what exists in the image rather than inventing.
Run the creative upscale
Topaz Bloom 2 on the same source, in parallel. It rebuilds the materials compression had erased, and that is where you look closely at the faces.
Compare with the original
The source photo stays on the canvas next to both renders. A feature that has moved, an ear that changed shape, a piece of jewellery that appeared: that is what decides which of the two you keep.
The same canvas, tuned for different needs: each variation is two or three lines of prompt away.
The direct use case: a shot from a 2000s compact, or a photo of a photo taken on a phone. Precision is almost always enough, and it is the right one here since likeness is all that matters. For a damaged print, scratched or yellowed, follow up with restoration.
A visual recovered at 600 pixels will not hold on a poster. A 4x takes it to 2400, which prints normally. Read the output size the node announces before generating: it tells you straight away whether you have enough.
Anything that travels through a messaging app comes back recompressed and shrunk. Enlarging recovers part of what the compression ate. On text that has become unreadable, however, neither upscaler will bring it back correctly.
This is Bloom 2's ground. Nobody knows the exact grain of a wall or the weave of a sofa, so nobody can hold it against the model for inventing them. The render is often more convincing than the source.
An image out of an Image node wires straight into an Enhance node. Useful when the generating model tops out in resolution and the deliverable needs bigger.
The Enhance node also offers video, with ByteDance Upscaler and Topaz Proteus. The source audio track is kept. This template only shows the image case, but the gesture is the same.
The mistake that ruins a family photo. The render is prettier, and that is exactly the trap: a redrawn jaw and an added earring go unnoticed on screen, but not by someone who knows the person.
An editing model handed three hundred pixels does not restore a face, it invents one. Enlarging afterwards only enlarges the invention. The right order is enlarge, then retouch.
An upscaler reveals detail drowned in noise, it does not create absent information. A backlit face, a distant number plate, text three pixels high: what comes out will be plausible and false.
A 4x followed by another 4x does not give you sixteen times better, it stacks the artefacts of the first pass. If the output is not enough, go back to the best available source rather than multiplying passes.
Precision as soon as there is a face someone has to recognise. Bloom 2 for a background, an object, a texture, anything whose exact detail nobody knows. When in doubt, generate both: the canvas puts them side by side and the answer is obvious.
No, and that is the honest limit here. An upscaler restores detail that is present but drowned in noise; it does not bring back a face lost to backlight or text too small to have been recorded. What Bloom 2 produces in those cases is plausible, not true.
After. An editing model handed three hundred pixels invents a face, and enlarging an invented face still leaves you an invented face. Enlarge first, retouch afterwards if the render calls for it.
Forty-two cents for both 4x enlargements. The price shows on each node before you click, and it depends on the OUTPUT size: enlarging a large image costs more than enlarging a thumbnail.
About a minute and a half against a few seconds, because it regenerates the material instead of interpolating it. That is also what makes it markedly more expensive per megapixel. The node announces the price before it runs.
No. Your media stays in your workspace, it is not handed to any training run. The question matters all the more here, since these are family photos.
Yes, the canvas works on mobile and the photo is taken straight into the Media node, so you can handle a print lying on a table. The creating on mobile guide covers the canvas gestures by finger.
Restoration repairs physical damage, scratches, creases, stains, and colorises. This one repairs nothing: it adds pixels. The two complement each other, and the order is enlarge first.
Sharpen a photo with AI
Use this workflowFree account, no card required. The workflow opens pre-filled in your canvas.
Topaz Precision
Flux Kontext