Start from the best source you have
This single decision affects the result more than any setting in the tool. An upscaler reconstructs detail that compression removed — it cannot recover detail that was never in the file you uploaded.
The version you emailed to yourself, downloaded from a chat app or saved off social media has already been recompressed and usually downscaled. Find the camera original or the first export. A sharp 1500px original beats a 3000px file that has been through three rounds of re-saving.
- 1920px and above — excellent material, plenty of real detail to build on.
- 1000–1900px — the most common starting point, reliable for photos.
- 500–1000px — usable, but check faces and text at full zoom.
- Under 500px — the model is inventing most of the image. Sometimes convincing, often not.
Pick the mode that matches your content
Different content breaks in different ways, which is why there is a content selector rather than one button that claims to handle everything.
Photographs
Skin, hair, fabric and foliage are all texture, and texture is what reconstruction does best. Leave the mode on General and let the model rebuild the fine detail that compression flattened.
Text, screenshots and interfaces
Choose Text & Screenshots. This is the one case where a generative model actively hurts: asked to enlarge a letter it cannot fully resolve, it will draw something that merely resembles the original character. The text-safe mode uses a deterministic upscaler that enlarges and sharpens without redrawing, so the letters stay the letters.
Anime, line art and logos
Choose Art, Logos & Pixel Art. Illustration is built from hard edges and flat colour regions, the opposite of photographic texture. A photo-tuned model adds halos to clean line work and invents grain inside areas the artist meant to be flat.
Old and damaged photographs
Scans need repair before enlargement — grain, colour shift and physical damage are separate problems from resolution. See restoring old photos for the full workflow, including how to scan so the software has something to work with.
Choose the factor by arithmetic, not ambition
Work out what you actually need: divide your target width by your current width. A 1920px image destined for a 4K display needs 2×, not 8×.
Reaching for the maximum factor by reflex makes things worse, not better. Every additional step of magnification is more of the picture being estimated rather than derived from your pixels, and that is where output starts looking artificial.
What a free tier realistically gives you
ImageUpg's free plan runs 7 tokens a week, refilled every Monday, and a 2× upscale costs one token. That is seven real upscales a week on the same engine the paid plans use for photo work — not a watermarked demo of a weaker model.
Creating an account is free and removes the watermark. Paid plans exist for the things a weekly allowance cannot cover: much higher factors, larger monthly pools, priority processing and the heavier reconstruction modes.
How to judge the result
Zoom to 100%, not fit-to-screen. Every upscaler looks impressive at 25% zoom, because at that size you are not seeing the pixels it invented.
Then look at the four places reconstruction fails first:
- Eyes and teeth — the first place invented detail becomes uncanny.
- Small text — either the glyphs are intact or they are not; there is no middle ground.
- Repeating patterns — fabric weave and brickwork expose models that smooth rather than rebuild.
- Flat areas — sky and painted walls should stay smooth, not acquire texture that was never there.
If those hold up, the rest of the image will.
What upscaling cannot fix
Worth knowing before you spend tokens on a lost cause:
- Motion blur — the detail was never captured, only movement.
- Severe out-of-focus — same reason. A model can sharpen an edge; it cannot infer a face the lens never resolved.
- Blown highlights — pure white has nothing underneath to recover.
- Tiny, heavily compressed files — a 200px image saved at low quality has lost its structure; rebuilding from that is invention rather than enhancement.
The rule of thumb: upscaling recovers detail that was compressed away, not detail that was never there.
Frequently asked questions
Only for anonymous use. Creating a free account removes the watermark, and the account itself is free. Always check the downloaded file rather than the on-screen preview.
Yes. ImageUpg gives anonymous visitors 1 token with no account at all. Creating a free account raises that to 3 tokens every week and removes the watermark.
For most photographs 4x is the practical ceiling for natural-looking results. Beyond that the model is inventing the majority of the pixels, so the output depends heavily on how much real detail survived in the source. A sharp 2000px photo upscaled 4x will beat a blurry 400px photo upscaled 8x every time.
ImageUpg processes images through its GPU pipeline and does not train models on user uploads. For any tool, if the privacy policy is vague about training data, assume the answer is unfavourable.
Denoising stages are usually the culprit. A denoiser that flatters skin texture will smooth the edges of letters until they read as blurry — which is why ImageUpg runs no denoise pass on its 2x mode, and routes faces through dedicated restoration models instead of a general-purpose smoother.