October 6, 2026
AI upscaling: faithful or creative, what an upscaler really invents
The same 360-pixel photo upscaled four times by classic interpolation, by Topaz Precision and by Topaz Bloom 2, cropped at 100%. What an AI upscaler actually does, when creative mode changes a face, when faithful is enough, and what each one costs.
Frank HoubreFounder of Imaginode
Upscaling means making up fifteen pixels out of sixteen
A 4x upscale multiplies the pixel count by sixteen: fifteen out of sixteen pixels did not exist in the source image, and the whole difference between two upscalers is how they make them.
Simple numbers first. A photo 360 pixels wide contains 129,600 pixels. Upscaled four times, it becomes 1440 pixels wide, or 2,073,600 pixels. The original information has not changed: there are still 129,600 real measurements of the scene, and almost two million pixels that had to be invented to fill the rest.
So the question is never "does the tool add detail". It has to, otherwise the image would stay blurry. The real question is where that detail comes from. From a calculation on neighboring pixels, from a model that has learned what a sharp texture looks like, or from a model that redraws the scene its own way. Those are three different jobs, and they give very different results on a photo of a person.
That is what we look at here, pixel by pixel, on a real photo, with the two image upscalers available on Imaginode and the classic enlargement of any photo editor as the baseline.
The three families of upscalers
Interpolation computes an average and gives blur, faithful super-resolution rebuilds plausible edges from the image, and creative upscaling regenerates texture with a diffusion model that may change what it sees.
The first family is interpolation: bilinear, bicubic, Lanczos. Each new pixel is a weighted average of its neighbors. It is fast, free, built into every editor, and perfectly honest: nothing is invented. The price of that honesty is blur. A sharp two-pixel edge becomes an eight-pixel gradient.
The second family is what we will call faithful super-resolution. A network was trained on millions of image pairs, the same scene in low and high definition, and it learned to rebuild crisp edges, eyes, hair, text. It sticks to the input image: its job is to guess the sharp version of what is already there. Topaz Precision belongs to this family.
The third family is more recent. Creative upscaling relies on a diffusion model, the same kind of model that generates images from text. The low-resolution image acts as a guide, and the model regenerates the texture on top. The result is often spectacularly sharp, with skin, fabric, material. But the model is not bound to respect what it sees, only to stay consistent with it. Topaz Bloom 2 belongs to that one.
The test: a 360-pixel family photo
The same blurry 360 × 360 photo, upscaled four times to 1440 × 1440 by Lanczos, by Topaz Precision and by Topaz Bloom 2; the crops below are at 100%, untouched.
The case is deliberately hard, because it is the most common real one: an old family photo, small, compressed, a bit blurry, with four faces. If an upscaler is going to betray something, it will be there, and that is where it matters most.
All three versions are exactly the same size, 1440 pixels wide. The crops show the same 560-pixel square around the father's face, without resizing, as you would see it zoomed to 100% in an editor. Click an image to enlarge it and move between them with the arrows.
The classic enlargement is what you would expect: the photo is bigger and not any sharper, with the JPEG compression blocks of the original now plainly visible. Topaz Precision gives a sharp face, readable eyes, three-day stubble, and it is clearly the same man. Topaz Bloom 2 gives the sharpest image of the three, and a different man.
What creative mode changed, detail by detail
On this photo, Bloom 2 changed the shape of the father's face, his teeth and his eyes, turned a t-shirt into a collared shirt, regularized the stripes on the child's sweater and added a hand on the mother's shoulder.
Put the two versions side by side and take inventory. Bloom's father has a wider face, different cheekbones, redone teeth and a gaze that is no longer his. The light t-shirt under his sweater has become a collared shirt. The boy's striped sweater, an irregular pattern in the original, is now a perfectly even knit straight out of a catalog. On the left edge of the frame, a blurry shape has become a hand resting on the mother's shoulder.
None of these changes is absurd. That is exactly the problem: each one is plausible, and nobody will notice except the family. The model saw the outline of a smiling man in his mid-thirties and drew a smiling man in his mid-thirties, with a lot of skill. It just is not the man in the photo.
Precision is not perfect either. Skin is smoothed, almost plastic in places, the girl's pink sweater keeps odd streaks where compression had crushed the information, and the background stays soft. But when it does not know, it leaves things blurry. It does not invent an object to fill the gap, and that is exactly what you want from a tool working on a real photo.
When to choose faithful upscaling
As soon as the image shows something real that has to be recognized, a person, a product, a document, a place, faithful upscaling is the only defensible choice.
The rule is easy to remember: if someone can say "that's not him" or "that's not our product", you want faithful. Family photos, client portraits, product shots with a label or a logo, architectural drawings, screenshots, archives, evidence. On all of these, an invented detail is a mistake, not an improvement.
It is also by far the cheapest. Topaz Precision is billed per 24-megapixel block of output: 12 credits for any image that comes out under 24 megapixels, which covers a 4x of a 1024-pixel image or a 2x of a 4-megapixel image. Per megapixel, it costs twelve times less than creative mode.
For an old damaged photo, with folds, stains or faded colors, upscaling alone is not enough: it enlarges the flaws too. The chain that works is described in restore an old photo, and for a photo that is simply small and blurry, sharpen a photo starts with the right settings.
When creative mode is worth the price
Creative mode is the right choice when the image has no truth to respect: an AI-generated image, an illustration, a landscape, a texture, a set, especially if it is going to be printed large.
An AI-generated image has no original. The castle, the forest, the invented character exist nowhere, and nobody can say the stone in the wall did not look like that. In that case, the freedom of creative mode becomes an asset: it adds material where the generator left smooth surfaces, and the image gains a lot at 100%.
Same logic for landscapes, architecture seen from afar, background textures, painted illustrations. Anything made of material rather than identity copes well with reinvented grain. It is often the difference between a poster that holds up close and a poster that looks stretched.
It comes at a price, though. Bloom 2 is billed per 2-megapixel block of output, twelve times more per megapixel than Precision: 34 credits for a 2x of a 1024-pixel image, 101 credits for a 4x of the same image, which comes out at 16.8 megapixels. It is also slow: 149 seconds measured on a small image during our tests. And even on a generated image, check the faces before you accept it: a recurring character can change heads from one image to the next.
What no upscaler can recover
A face thirty pixels tall, heavy motion blur, missed focus or unreadable text cannot be recovered: faithful mode leaves them blurry, creative mode replaces them with something else.
There is a floor of information below which no tool can go. A face thirty pixels tall in the source does not hold enough measurements to recover a person. Faithful mode will produce a smooth, vague face, creative mode a precise, wrong one. Neither will produce the right one.
Text is the most treacherous case. A sign, a plate, a label that is unreadable in the source will either stay unreadable or be rewritten with plausible letters that mean nothing, or worse, that mean something else. If the text matters, retype it on top instead of counting on the upscaler. The same problem shows up at generation time, covered in AI hands and text problems.
Last case, the frame itself. An upscaler makes the image bigger, it does not make the field of view wider. If you are missing scenery on the sides to go from a square to a poster format, that is another tool: extend an image adds consistent scenery at the edges, and the upscale comes after.
2x or 4x: the right factor depends on the mode
Output size is the input size multiplied by the square of the factor: a 4x produces sixteen times more pixels, which barely changes the price in faithful mode and multiplies it by three or four in creative mode.
The calculation to do before you click: width × factor, height × factor, and the product in megapixels. A 1024 × 1024 image comes out at 2048 × 2048 in 2x, or 4.2 megapixels, and at 4096 × 4096 in 4x, or 16.8 megapixels. A 2048 × 2048 image comes out at 16.8 megapixels in 2x and 67 megapixels in 4x.
In faithful mode, the 24-megapixel block makes 4x almost free on small images: 12 credits in 2x as in 4x for a 1024-pixel image. You might as well take 4x if you are unsure about the final use. In creative mode it is the opposite: 34 credits in 2x, 101 in 4x for the same image. Take the factor you need, not one more.
A ceiling applies in both cases: 100 megapixels of output. Beyond that, the file gets too heavy to handle comfortably, and the creative price becomes unreasonable. A 4096 × 4096 image therefore cannot go 4x, which would give 268 megapixels. To know how many pixels you really need for a given medium, the math is in printing an AI image large.
Upscale last, never in the middle
Upscaling is the last step of a chain: generate, retouch, approve, then upscale the final version, because a model placed after the upscale does not benefit from its pixels.
The classic mistake is upscaling too early. You generate an image, upscale it because it looks great, then retouch it or send it to a video model as a start frame. The pixels you gained are lost: an image model reworks at its own resolution, and a video model outputs 720p or 1080p whatever the size of the image you feed it.
The right order never changes. Explore at the cheapest tier, pick, retouch, approve the composition, then upscale once, the version that actually goes into production. It is the same principle as for video, where you approve in low definition before scaling up, covered in AI video in 4K.
On the canvas, you can read it directly: the Enhance node plugs in at the end of the chain, on the output of the last Image node. If you change something upstream, you rerun the chain and the upscale follows. The building blocks are explained in what is an AI workflow.
Where to upscale on Imaginode
Three places: the canvas Enhance node, in Image, Video or Sound mode; the upscale action on each layer of the Retouch node; and the edit_image tool of the MCP server, from Claude or ChatGPT.
The Enhance node is the main entry point. In Image mode, it offers Topaz Precision and Topaz Bloom 2, with a 2x or 4x factor and the price shown on the button before you run it, computed on the actual size of your image. In Video mode, it offers the two video upscalers, covered in the best AI video upscalers.
In the Retouch node, each image layer has its own upscale action, at the same price. That is handy when you build a poster from several elements: you upscale the element that needs it, not the whole document. And if you work from an assistant, the edit_image tool of the MCP server upscales an image without opening the canvas.
To compare with what exists elsewhere, Topaz Gigapixel, Magnific, Krea, Upscayl and the rest, the full comparison is in the best AI image upscalers.
The summary
Faithful for everything real, creative for everything invented, the factor chosen according to the mode, and the upscale as the very last step.
A person, a product, a document, an archive: Topaz Precision, 12 credits up to 24 megapixels of output, and 4x costs nothing more on a 1024-pixel image. The result can stay a little smooth, but it will not lie.
A generated image, an illustration, a set, a texture, especially for print: Topaz Bloom 2, billed per 2-megapixel block, so use it sparingly. Check faces at 100% before you accept the result, and prefer 2x when 4x is not needed.
And in every case, remember that an upscaler is not a miracle worker: below thirty pixels of face or inside motion blur, the information is gone. Better to know it before promising someone you will "recover" their photo.