Commons:Upscaling
 |
The proposal may still be in development, under discussion, or in the process of gathering consensus for adoption. References or links to this page should not describe it as "policy". |
- incorrectly interpreted some photo damage (eg. the black marks, one of which is very close to her hair) as a detail of the subject
- interpreted ambiguous details as damage and removed them (the left upscale removes what may be a ring and what may be a thin bracelet; the right upscale removes what may be a second bracelet);
- made different and speculative interpretations of ambiguous aspects of the photo (eg. the material and width of her bracelets; the fabric of her dress; the shape of her nose; the colour of her eyes; the direction she is looking in)
Upscaling is the use of software to increase the pixel count and size of an image. In general, images on Commons should not be upscaled, and the use of artificial intelligence (AI)-based upscaling tools in particular should be avoided. Exceptions are made for images in use on Wikimedia projects and where the original image is inaccessible.
Uses of upscaling
Upscaling is used for many reasons. For example, people who want to print images may get better results with an upscaled version. Images displayed on the web are often low resolution by default, but at times a very small image that would otherwise appear heavily pixelated is upscaled for visibility. Upscaling images with text can similarly render the text more legible.
Commons regards upscaling for printing or personal usage to be up to end users. In other words, we do not upscale images in anticipation of users' needs, but rather provide originals that users can modify however they like.
An exception to these rules is where there is demonstrable consensus on a Wikimedia project to upscale a specific image for use on that project, in which case Commons regards the file to be in scope. Another, less common, exception is for upscaled images where the upscaled version is the only accessible version.
Note that upscaled photographs of identifiable people must also comply with Commons rules for photographs of identifiable people and, if AI-based software was used, the guideline for AI images of identifiable people.
Software
Many software products enable upscaling, but use different means to accomplish a higher resolution. The traditional method (used by Adobe Photoshop, for example), samples pixels within the original image. Newer software often uses artificial intelligence, inferring edges, faces, objects, and other details based on not just the original image but also images in the model's training data.
The quality of AI-based upscaling varies from tool to tool. Uploading an image to popular chatbots, for example, will often result in other changes in ways that are unacceptable for Commons, even if the AI is only instructed to resize the image or repair minor damage to it. Other models like Topaz Gigapixel, advertised as "improving image resolution", can produce a result that looks more like the original, but may still introduce details that did not exist in the original.
Issues with AI tools
When AI upscaling increases the resolution of a low quality image, it cannot accurately restore the information that is missing. It can only provide a plausible, speculative interpretation of what the patterns of light and shade could have originally represented. If the same image is upscaled by different AI services (or even the same service at different times), this will produce different results. None of them will be the "true" original image.
When upscaling removes grain or damage from a photo, the AI again can only offer likely possibilities of what was underneath it. When removing a scratch in a clear blue sky, this will probably be harmless, but when the damage covers some detail of a photograph, the AI will have to generate original, speculative content to replace it.
AI upscaling may not always correctly factor in the wider context of the image. A photograph from a particular decade and country may be processed in a generic way based on the average of the AI's training data. Portraits can tend towards modern American aesthetics around makeup, teeth and facial expressions. 19th century photos of cities can have anachronistic details added, such as flagpoles being turned into TV aerials.
Handling upscaled images
While most upscaled images are not appropriate for Commons, exceptions for files in use on other Wikimedia projects must be handled properly.
- All upscaled images on Commons must use the appropriate template on the file page. For upscaling using AI-based tools, use {{AI upscaled}}; for others, use {{Retouched}} with a description that mentions upscaling.
- In addition to the templates above, ideally the description of upscaled images should also be clear that it has been upscaled, and should include the name of the software used, if known.
- If the original image is accessible, it should also be hosted on Commons, separately from the upscaled version, and prominently linked from the upscaled version.
Incidental expansion
While upscaling of an image is generally not permitted on Commons, certain editing techniques may expand parts of an image in ways that are technically similar to upscaling, but are not considered problematic and not subject to this guideline.
- Perspective correction commonly involves a portion of the image being expanded. For example, it may not be possible to photograph a painting straight-on, either due to the height at which it is hung, reflections, or other technical issues. The only way to produce a good image is to photograph from some other angle and adjust perspective. Similarly, photographs of architecture are often adjusted to straighten vertical and/or horizontal lines.
If you are correcting the aspect ratio of an image, it is equally permissible to upscale one dimension as to downscale the other.
In both of these cases, there is a choice here between loss of information in one dimension (by shrinking) or adding interpolated, possibly slightly incorrect, information in the other (by stretching). Neither of these is considered more desirable than the other, but such photo edits should be indicated clearly either by using {{Retouched}} or a similar template, and/or through the text of an edit summary for an overwrite.