Different photos fail in different ways, and the right upscaling settings depend on why the image looks bad in the first place — a blurry portrait, an under-sized product shot, and a scanned family photo each need a different approach. This guide walks through common use cases and how to get the most natural result from each. If you want the underlying theory first, see how AI image upscaling works.
Faces are the hardest subject for any upscaler to get right, because human vision is extremely sensitive to anything that looks slightly "off" about skin, eyes, or hair — long before it can articulate why. For portraits:
- Use Ultra Mode when the photo is the main subject of the output (a printed portrait, a professional headshot, a keepsake photo). It combines face restoration with background enhancement, so the subject doesn't end up sharper than their surroundings, which is a common giveaway of AI processing.
- Use HD Fast for quick previews or profile pictures, where processing speed matters more than pixel-level fidelity.
- Watch for over-smoothing on skin. If a result looks slightly waxy, it's usually because the source photo was already soft or low-resolution to begin with — the model has less real detail to work with. Starting from the sharpest available version of the photo (not a re-saved copy or a screenshot of it) makes a measurable difference.
Product photos have different priorities: clean edges, accurate texture and material rendering, and consistency across a catalog.
- 4K Pro is usually the right default for product images — it balances enhanced detail with speed, which matters when you're processing a full catalog rather than a single hero image.
- Batch processing is built for this. Pro and Business plan users can upload up to 20 images at once for parallel processing at the same quality setting, which keeps a whole product line visually consistent instead of upscaling images one at a time with slightly different results.
- Check fine text and logos closely. Small printed text or brand marks on packaging are exactly the kind of "novel" detail that upscaling models are more conservative about reconstructing — verify legibility on the result before publishing, rather than assuming it upscaled cleanly.
- Upscale before, not after, other edits. Background removal, color correction, and cropping are easier to apply cleanly to a properly upscaled image than the other way around, since upscaling after heavy editing can amplify compression artifacts introduced by earlier processing steps.
Restoring old family photos is one of the most rewarding uses of AI upscaling, but also one where expectations need to be realistic.
- Scan at the highest resolution your scanner supports before uploading. A high-resolution scan of a small, low-quality original print still gives the model more real information to work with than a low-resolution scan of the same print.
- HD Fast's face restoration works well on portraits from old prints, where the main goal is making faces clearer rather than reconstructing fine background detail.
- Physical damage (creases, tears, discoloration) is a different problem from low resolution. Upscaling will not repair torn or heavily faded photos — it increases resolution and sharpens detail that's present, but it isn't a restoration tool for physical damage. Photos with heavy damage will show that damage more clearly at higher resolution, not less.
- Sepia or faded color photos will upscale in their original color profile. Upscaling doesn't correct color fading — pair it with separate color correction if you want both a sharper and more vibrant result.
- HD Fast is almost always the right choice here — these images rarely need print-level fidelity, and fast turnaround matters more when you're preparing content for social posting or a presentation.
- Screenshots of screenshots compound quality loss. If you're upscaling a screenshot, use the most original version available — each re-save or re-share (especially through messaging apps) recompresses the image and removes detail permanently.
- For graphics with text or UI elements, check the result carefully. Upscaling models are trained primarily on photographs; flat graphics, icons, and rendered text can occasionally upscale less predictably than photographic content, since they don't follow the same texture statistics the model learned from.
- Format matters less than you'd think. We support PNG, JPG, JPEG, WebP, and HEIC as input, with PNG output to preserve maximum quality after processing — but a low-quality JPEG will still cap the result at the detail actually present in the file, regardless of format.
- File size limit is 5MB per image — if you're working from a much larger original (e.g., a RAW camera file), export a high-quality JPEG or PNG version first rather than a heavily compressed one.
- Use the before/after slider before downloading. It's the fastest way to catch whether a specific region (eyes, text, fine patterns) upscaled the way you expected, rather than judging the whole image from a quick glance at the thumbnail.
- When in doubt, try HD Fast first. It's free and fast, so it's a reasonable way to judge whether an image is a good candidate for upscaling before spending credits on 4K Pro or Ultra Mode for the same file.
Matching the tier and workflow to what the photo actually needs — rather than always reaching for the highest setting — is usually what separates a natural-looking result from one that looks obviously processed.