The verdicts, re-tagging, enlarging and working on images inside PDFs.
A customer sends a picture and somebody has to decide, before the job goes on press, whether it will print. Image Rescue answers that in a sentence, and then does whatever is worth doing about it.
It is deliberately not a magic button. Of the files that arrive labelled “low resolution”, a large share are perfectly good and only tagged wrong, and a smaller share cannot be saved by any amount of processing. Telling those two apart is worth more than the enlarging is.
Effective dpi: the pixels in the file divided by the inches it occupies on the finished piece. Nothing else is a resolution.
| The file says | What is actually true |
|---|---|
| 3000 × 2000 pixels, tagged 96 dpi, placed 10″ wide | 300 dpi. Perfect. Only the tag is wrong — re-tagging fixes it and does not touch a pixel. This is the commonest case there is. |
| 1200 × 800 pixels, tagged 300 dpi, placed 8″ wide | 150 dpi. The tag says 300 and the tag is irrelevant: placed at twice its natural size it is half the resolution. |
| 960 × 640 pixels, tagged 96 dpi, no placed size given | 96 dpi at 10″ wide — because with nothing else stated, the tag is what decides the size it lands at. |
| Verdict | What it means and what to do |
|---|---|
| Already press ready | Enough pixels at the size it is placed. Nothing to do. |
| Only the dpi tag is wrong | The pixels are right; the tag is not. Re-tagging is byte-for-byte lossless — the image data is not touched at all. |
| Can be enlarged | Within what this kind of content carries. Go ahead. |
| Will be soft | Past the comfortable ceiling. It will print, and it will look soft next to the rest of the job. Offered as an override, not as the obvious next click. |
| Too small to save | Nothing recovers detail that was never captured. The screen gives you the exact pixel size to ask the customer for. |
| Ask for vector artwork | It is a logo held as pixels. It will enlarge better than a photograph would, but the right answer is the AI, EPS or PDF from whoever drew it — that prints perfectly at any size and ends the conversation permanently. |
The ceiling depends on the content, because content behaves differently. Hard-edged line art reconstructs well and carries about 4×. A photograph carries about 2×; past that, sharpening only outlines the blur. Those numbers are what stop this utility making promises it cannot keep.
Three things, in this order, because the order is most of the quality:
| Step | Why |
|---|---|
| Clean | Only when the measurements say there is something to clean. JPEG compression works in 8-pixel tiles, and enlarging first turns each one into a visible block; cleaning first removes them while they are still one pixel wide. The filter compares neighbours by how similar they look, so a real edge survives and mottle in a flat sky does not. |
| Enlarge | Lanczos, in linear light. sRGB values are not proportional to light, and averaging them directly is what makes a naive enlargement look muddy. Line art uses a filter that does not ring instead, because a dark fringe beside a logo is what customers ring up about; a screenshot enlarged by a whole number uses nearest neighbour, because every soft filter turns crisp UI text to mush. |
| Sharpen | With a threshold, so flat areas are left alone. Sharpening grain and sky is what makes an enlargement look worse than the file it came from. The amount comes down as the enlargement goes up, because past about 2.5× there is no fine detail left to sharpen. |
Everything it writes is tagged with the dpi it was made for, so it places at the right size in InDesign, Illustrator or a RIP without anyone scaling it by hand.
The Suite AI (Settings → AI) is a text model. It never sees the image and never touches a pixel — it cannot; that is not what a text model does.
What it is given is the measurements: pixel size, effective dpi at the placed size, how blocky the JPEG is, how much of the frame is flat, how many colors there are, how much grain sits in the smooth areas. What it is asked for is the judgement an experienced prepress operator would make from the same numbers — which treatment, how hard, and what to say to the customer. It can also draft the “please send us the original” email, with the exact pixel size in it.
Three things keep that honest. There is always a measured answer first, and it is a good one; if the AI is off, unreachable or answers with nonsense, that answer stands and the screen says which you are looking at. Every number the model returns is clamped before it can reach the resampler. And the model cannot overrule the verdict — it may not turn “too small to save” into “enlarge it”, because that ceiling is arithmetic, not opinion.
A PDF is the best thing you can hand this utility, for one reason: the file already knows how big every picture in it is being printed. A loose JPEG has to be asked what size it is placed at. A PDF states it, in the placement matrix, for every image on every page. So the resolution figures for a PDF are exact rather than assumed.
Drop one in and you get a row per image, worst first: where it is, how many pixels it holds, how many inches it covers, and therefore its real resolution. An image used more than once — a logo repeated across a sheet — is judged at its largest use, because that is the one somebody will notice.
There is no “re-tag” verdict for a PDF, because there is no dpi tag to be wrong about. Inside a PDF the placement is the resolution.
| Left alone | Why |
|---|---|
| CMYK images | The one refusal worth arguing about, and it stands. Reading CMYK out and writing RGB back re-separates the customer’s artwork: a black built from four inks comes back as a muddy four-color black, neutral grays go color-cast, and nobody spots it until it is on press. Enlarge it outside the PDF and have it re-placed by whoever built the file. |
| JPEG 2000, JBIG2, CCITT fax | Compressions this suite carries no decoder for. CCITT in particular is 1-bit scanned line work, which upsampling cannot improve anyway. |
| Indexed, Separation and DeviceN | These carry palette or ink meaning that a straight RGB rewrite would throw away. |
| Stencil masks and images with a Decode array | A 1-bit mask is drawn at the resolution of the press, not of the file. A Decode array deliberately remaps the samples, and a rewrite would lose that. |
| Inline images | These live inside the page content stream itself rather than as their own objects, so replacing one means rewriting the page — exactly what this refuses to do. They are counted and reported so you know they are there. |
Every one of those is reported with its resolution anyway. Knowing that a CMYK photograph is 90 dpi at full bleed is worth having — especially when the answer is a phone call to the designer rather than a button here.
Transparency is carried through: an image with a soft mask is enlarged along with its mask, so a knocked-out picture keeps its edge. Photographs that arrived as JPEG go back as JPEG, since the compression artefacts are already baked in and a lossless rewrite would multiply the file size for nothing; flat artwork and type go back lossless, because JPEG rings around every hard edge. A rescued PDF is bigger than the one it came from — more pixels is a bigger file, and that is the trade.
A trained super-resolution model genuinely can invent plausible detail, which on a photograph past 2× is the difference between soft and acceptable. Rather than bundle several hundred megabytes of weights, Image Rescue runs whatever this shop already owns. Point Settings → AI → External upscaler at Topaz, Real-ESRGAN, Upscayl or anything else with a command line, and test it there on a small pattern.
When one is configured, the model does the enlarging and the built-in filter finishes to the exact pixel count the press target asks for — a tool that only does whole-number scales leaves you at 4× when 3.1× was wanted. If it is missing, misconfigured or fails, the built-in filter does the whole job and the report says so rather than failing the file. Having one also raises the honest ceiling, which is why a file the shop would otherwise refuse may be offered when the model is available.