What a re-encode keeps and what it throws away
An image file is pixels plus a set of claims about them: which colour space those numbers are measured in, whether there is an alpha channel, what the camera recorded in EXIF — model, exposure, date, often GPS coordinates — and, in GIF, WebP and AVIF, whether there is more than one frame. Almost every operation here decodes the file and writes a new one, and each of those claims has its own fate.
Lossy and lossless is the first fork. PNG and BMP store the exact pixels, so re-saving them a hundred times changes nothing. JPG, AVIF and ordinary WebP discard detail the eye is bad at noticing, and the discarded part is gone — converting a JPG to PNG cannot restore it, it only stops the bleeding. That is why re-saving a JPEG twice is worse than saving it once: the second encode quantises pixels that are already quantised, and if the picture was cropped or resized in between, the new 8×8 block grid no longer lines up with the old one, so the errors compound instead of repeating. Work from the original whenever you still have it.
Chroma subsampling is the specific reason a JPG screenshot looks wrong. JPEG normally stores brightness at full resolution and colour at half, which is invisible on a photograph and obvious on red text or a thin coloured line. If your image is a screenshot, a logo or a diagram, PNG or WebP is the right target and JPG is not.
Transparency has to go somewhere. JPG has no alpha channel at all, so PNG to JPG and AVIF to JPG ask you which colour replaces the transparent pixels rather than flattening them to white — or to black — behind your back. PNG, WebP and AVIF all carry full alpha, so PNG to WebP and the cut-outs from Remove Background keep their soft edges intact.
EXIF does not survive. Every tool that re-encodes — convert, compress, resize, rotate, crop — writes a clean file with no EXIF block, so camera model, timestamp and GPS location are dropped. Before publishing a photo that is exactly what you want. Orientation is the one piece of EXIF that matters visually, and Rotate Image sidesteps it by moving the actual pixels, so the result is upright in viewers that ignore the flag as well as in those that honour it.
Colour profiles are converted, not carried. Your browser decodes a Display P3 or Adobe RGB file into sRGB in order to draw it, and the file written back has no profile embedded. On an ordinary screen nothing changes; on a wide-gamut screen the most saturated colours have been clipped. TIFF to JPG goes through the same path, which is why a CMYK studio TIFF comes out as RGB and should stay in prepress software if the colours are being proofed.
Animation becomes a still. Animated GIF, WebP and AVIF files convert as their first frame. That is usually what a person wants from GIF to PNG, and it is worth knowing before you drop a twelve-second loop on it.
Where a browser is genuinely the wrong tool
Running in the browser is what makes these private, and it costs something. We would rather say so here than have you discover it halfway through a job.
Very large images. Every pixel is held in your device's memory, decompressed — a 100-megapixel image is around 400 MB as raw RGBA before anything else happens. Somewhere past that the tab runs out of room, and a 4× upscale hits the ceiling sooner still, at roughly a 6000×6000 output on a typical laptop. A desktop tool that streams tiles from disk has no such limit.
Camera RAW. CR2, NEF, ARW and DNG files are sensor data plus a development recipe, not finished pictures, and nothing here opens them. Export a JPEG or TIFF from your camera software first, or use darktable or RawTherapee — both free, both far better at the job than any web page could be.
Print colour. Spot colours, CMYK separations, ICC rendering intents and soft-proofing are the domain of prepress software. We convert colour into sRGB so the browser can draw it; we do not manage it.
Batch and automation. These tools run when a person clicks. Resizing forty thousand product photos on a schedule is a job for a command-line tool on a server, and the free ones are excellent: ImageMagick for almost anything, libvips for speed and low memory on huge files, mozjpeg and oxipng for squeezing the last bytes out, ffmpeg for anything with frames in it.
Background removal has a real ceiling. The model that runs here is 40–85 MB and downloads once — genuinely good on a clear subject, hair and fur included, and genuinely confused by a cluttered scene with no obvious subject. No amount of local compute fixes that; it is a limit of the segmentation model, and the paid services have the same one.
Choosing between the tools that sound alike
Compress Image vs Resize Image. Compress keeps the dimensions and spends fewer bytes per pixel. Resize changes the dimensions, which removes pixels altogether. For a 4000px photo that needs to be 200 KB, resizing first and compressing second beats compressing alone every time — and resizing is lossless in the sense that matters, because the pixels you keep are properly resampled rather than degraded.
Resize Image vs Upscale Image. Both change dimensions, and they are tuned in opposite directions. Resize is for making things smaller, with a Lanczos filter so detail survives the shrink. Upscale is for 2× and 4× enlargement, adding an edge-aware sharpening pass afterwards because enlargement always softens. Neither invents detail — what comes out was in your image.
Convert to ICO vs Favicon Generator. Convert to ICO produces one .ico file containing 16, 32, 48 and 256px versions, which is what a Windows application or shortcut wants. Favicon Generator produces the whole modern set — the .ico, the individual PNGs, the Apple touch icon, a web manifest and the HTML tags — zipped, which is what a website wants.
PNG to WebP vs JPG to AVIF. Both make web images smaller than the original, and they suit different sources. WebP keeps transparency and encodes in a moment, so it is the right target for screenshots, logos and UI graphics. AVIF compresses photographs harder — typically 40–60 per cent smaller than an equivalent JPG — but AV1 encoding is heavy, so expect a few seconds per large photo.
Blur Image vs Remove Background. Blur destroys information in a region you draw: faces, plates, account numbers. Remove Background separates subject from background and deletes the background entirely. If the goal is redaction rather than a cut-out, blur — and use coarse pixelation on text, because weak pixelation has been reversed in published research.