Commons talk:Upscaling
Add topicUpscaling discussions
[edit]Village pump
[edit]- 2020-03: Commons:Village_pump/Proposals/Archive/2020/03#Commons:Overwriting_existing_files_update, successful proposal to prohibit overwriting existing files with upscaled versions
- 2026-02 (in progress): Commons:Village_pump#More_explicit_policy_against_upscaling_needed
Other policies/guidelines
[edit]- 2023-10: Commons_talk:AI-generated_media/Archive_2#Custom_template_for_upscaled_images, developing a template for images upscaled with AI
- 2024-02: Commons_talk:AI-generated_media/Archive_4#AI_upscaling_of_historical_paintings, upscaling historical paintings
- 2025-12: Commons_talk:Featured_picture_candidates/Archive_28#AI-based_processing, FPC and AI-based processing
DRs
[edit]- 2024-04 (deleted): Commons:Deletion requests/File:Bernardetto de Medici - Giorgio Vasari.jpg, AI upscaled photo of a painting
- 2024-12 (kept): Commons:Deletion requests/File:(Venice) Fra'Mauro's World Map - Biblioteca Nazionale Marciana.jpg, AI upscaling, kept as in use
- 2026-01 (in progress): Commons:Deletion_requests/File:Alex_Pretti_Official_Portrait_High_Quality.jpg, upscaling altered the image (see also Snopes)
- 2026-02 (in progress): Commons:Deletion requests/File:Francis Marion Beynon.png, upscaling altered the image
- 2026-02 (deleted): Commons:Deletion requests/File:Pradip Baruah at his autobiography release event in Guwahati.png, upscaling altered the image
Serious doubts about the nutshell statement
[edit]All upscaled images must be properly labeled, with a link to the original, which should also be hosted on Commons.
So is this saying that if (for example) there is a useful image on Flickr, free-licensed but showing signs of upscaling, we cannot host that image on Commons unless we can contact the photographer and convince them to make the original available to us? - Jmabel ! talk 20:32, 12 March 2026 (UTC)
- Unusual case, but fair point. Maybe the language (there and elsewhere) should include something like "...if the original is accessible". — Rhododendrites talk | 20:41, 12 March 2026 (UTC)
- [1] — Rhododendrites talk | 20:49, 12 March 2026 (UTC)
- There is a difference between what is preferred/"best practice" and what is required. Throughout, we need to be clear which we are describing. - Jmabel ! talk 21:33, 12 March 2026 (UTC)
- Yep. Good point. Trying to smooth it out. As a guideline, rather than a policy, I think the main point here is being clear that (a) upscaling should be avoided by default, and (b) when an exception does happen, here are some rules for documentation. — Rhododendrites talk | 15:27, 14 March 2026 (UTC)
- There is a difference between what is preferred/"best practice" and what is required. Throughout, we need to be clear which we are describing. - Jmabel ! talk 21:33, 12 March 2026 (UTC)
- Noting that Commons:AI images of identifiable people currently only requires the original of an upscale to be stated
in the altered image's {{Information}} template
, not to be hosted on Commons. - There's an older thread open on the question at Commons talk:AI-generated media#Requiring source photos for AI-retouched historical images. As I say there, there are a lot of files in Category:Upscaling that have upscaled an old newsprint photo or a YouTube still without also uploading the original. (There was a spate, possibly still ongoing, of editors uploading upscaled YouTube stills of Korean pop stars and then outragedly denying that they had done any kind of AI processing to the images, perhaps because they were using an extraction tool that didn't describe it as "AI".) The English Wikipedia currently has 208 flagged articles that contain an AI upscaled image where the original (probably) isn't available on Commons to be simply swapped in. Belbury (talk) 09:09, 13 March 2026 (UTC)
- Thanks for flagging that discussion -- I hadn't seen it yet. I've modified the language here to be a bit softer (
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.
) — Rhododendrites talk | 15:32, 14 March 2026 (UTC)
- Thanks for flagging that discussion -- I hadn't seen it yet. I've modified the language here to be a bit softer (
Questions
[edit]Just want a clarification since it isn't specifically stated in the page. For files not in-use, is this draft guideline simply discourages upscaled files, or is it completely disallowing upscaled files? If it is the latter case, I think it should say "images on Commons must not be upscaled" instead in the first paragraph.
Also, in the "Handling upscaled images" section, it requires the description to state the image has been upscaled, and also requires the page to have {{AI upscaled}} or {{Retouched}} with a description. Is both of the requirements required, or just one requirement is enough?
I think requiring both might be redundant, and I think simply tagging the file with the templates should be sufficient. Thanks. Tvpuppy (talk) 03:59, 13 March 2026 (UTC)
- Templates are definitely preferable from the perspective of the AI Cleanup WikiProject on the English Wikipedia, as it means that the images get put into categories, and can be automatically scanned to see when they're being used (or misused) in Wikipedia articles (en:Wikipedia:WikiProject AI Cleanup/VWF bot log). Belbury (talk) 09:41, 13 March 2026 (UTC)
- I suspect there are just too many exceptions to make "must not be upscaled" practical. Worth a discussion if folks disagree, though.
- On the templates vs. description, I've modified it to prioritize the template and changed the language of the description to be softer. — Rhododendrites talk | 15:35, 14 March 2026 (UTC)
Example image
[edit]We may want to use a before-and-after example that better reflects how upscaling tends to get used on Commons, and is sympathetic to the idea that the upscaling user will often look back at the original photo and genuinely believe in good faith that they've made an improvement, or even been able to magically restore the "true" image. I'm not sure what the story is behind the Alex Pretti photo, but it's hard to imagine a user looking back at a side-by-side comparison and feeling that they'd simply made a higher resolution version of the original image. And at thumbnail size in the essay, it looks more like a photo edit than an upscale.
I haven't got a great example to suggest here, but here's another I encountered recently of an old newsprint photo being upscaled in a Wikipedia article (although in this case the user had found the upscaled version on on Pinterest and assumed it was a real hi res photo; I uploaded the newsprint afterwards). Belbury (talk) 17:11, 14 March 2026 (UTC)
- Good example. - Jmabel ! talk 19:03, 14 March 2026 (UTC)
- Another example was posted just yonderyesterday on the village pump, second image on the right side now, it shows a certain Mr. Buck. The Moiré pattern on the original image really appears deteriorating or actually destructive, but the upscaled image shows an idealized, statuesque face.

removal of Moiré effect results in hyperrealism - Another question, on the topic of "incidental expansion": some years ago I uploaded a couple of images of old oil portraits that hung high overhead and were difficult to capture (also, just using a rather bad phone camera I had on me). I rectified the images using gimp before upload, and the central top row image from here turned out like this (different source files). Today, I added the "retouched" template after reading the policy page. This is what was meant here? --Enyavar (talk) 13:32, 25 March 2026 (UTC)
- @Enyavar: On the latter: yes, or it could just indicate what it was derived from, but there should be some indication of how far it is from being simply photography, and now there is.
- On Mr. Buck: looks to me like a bit more conjecture than in the Umm Kulthum case. It may be as likely as any one interpretation of that very low-quality image, but it is definitely an interpretation. - Jmabel ! talk 22:18, 25 March 2026 (UTC)
Per the "welcome to assist in its construction" header on this essay I've swapped that Umm Kulthum example in, as well as adding an additional second upscale to make the convincing demonstration that different upscaling bots will give different results.
I've also replaced the mention of inference issues with a more expansive "Issues with upscaling" section. --Belbury (talk) 17:27, 22 April 2026 (UTC)
- Thanks, Belbury. Useful addition. I readded the Pretti, since there's room for both, and because it was a pretty notable instance of the problem we're talking about. Fair to say the "issues" section as you've written it should be retitled to be clear you're talking about AI upscaling in particular? — Rhododendrites talk | 18:06, 22 April 2026 (UTC)
- Thanks, and yes, I was just thinking the same thing about the section heading.
- I'd say the Pretti image was more illustrative of deliberate AI beautifying than AI upscaling - I think the essay gets off on slightly the wrong foot if the big "don't do this!" red warning icon is for a thing that people aren't really doing on Commons. Despite the Snopes source mentioning upscaling, as I understand it we don't know where it came from or how it was made, it was just going around on social media? Belbury (talk) 18:31, 22 April 2026 (UTC)
- Is it true it originated on social media. The amount of space given to the files on Commons in the Snopes piece gives the impression Commons was the misinformation vector. — Rhododendrites talk | 20:24, 22 April 2026 (UTC)
- The two Commons images in the Snopes article are (and always were) sourced as having been taken from Reddit and from LinkedIn. Belbury (talk) 20:48, 22 April 2026 (UTC)
- Is it true it originated on social media. The amount of space given to the files on Commons in the Snopes piece gives the impression Commons was the misinformation vector. — Rhododendrites talk | 20:24, 22 April 2026 (UTC)
Sharpening
[edit]I believe this policy should tackle both upscaling and sharpening as they are really similar processes, in fact you could bypass a restriction on generative AI upscaling by manually upscaling a low resolution picture by traditional methods like bicubic or bilinear interpolation and then using generative AI sharpening. It's moon (talk) 15:11, 6 April 2026 (UTC)
- IMO they are separate issues. Sharpening is involved in a lot of standard digital image post-processing. The AI versions of both upscaling and sharpening have some overlapping problems, but that's probably more for the more general AI content guideline. Upscaling is something that has a lot of specific applications, and something I think we can deal with pretty directly with a broad discouragement that wouldn't apply to sharpening. — Rhododendrites talk | 18:04, 22 April 2026 (UTC)
Proposal active
[edit]I've opened a proposal to make this a guideline here: Commons:Village_pump/Proposals#Survey (upscaling). Pings for people who have edited this page: @Belbury, Jmabel, It's moon, Enyavar, Tvpuppy, Jurta, and Grand-Duc: — Rhododendrites talk | 19:51, 23 April 2026 (UTC)
- @Rhododendrites Hi, do you have any thoughts on using AI to upscale maps? RiadS99 (talk) 17:26, 3 May 2026 (UTC)