Quantity Based: One of the continual problems the AI art generation faces is in quantity, though it is continually improving. For instance, in the past, AI art would struggle with getting the correct number of fingers correct, or perhaps the correct placement of knuckles and joints in the fingers.
General Softness & Low Resolution: AI art takes immense computing power to generate, and it still hasn’t streamlined this problem. So often, AI art is limited in resolution and detail.
Repetition: To further expand on the tip above, AI art often uses repetition to help speed up the generation process. So you may see something copied several times over the same image.
Asymmetry: Asymmetry exists in all facets of life, [… if you] photograph the building so that it looks symmetrical across the plane. AI doesn’t understand these rules and often creates subtle symmetry shifts in its images.
There are some indications that an image is created by AI showing wrong details of the human hand like 6 fingers. So far AI does not understand the semantic meaning of “hand” having only the visual demarcation of hands in images as trained by mechanical turks. Images of hands however, can be misleading for the trained eye where also good painters have difficulties.
Let’s have a closer look at the images of Princess Kate and their kids Charlotte (8, right) , Louis (5, left) and George, 10 (behind) by a check list that I developed earlier with another family member, the Andrew/Maxwell/Giuffre image that even fooled me in the beginning.
Image source: credible /dpa.
File Modification Date/Time : 2024:03:11 07:00:09+01:00
File Access Date/Time : 2024:03:11 16:27:05+01:00
File Inode Change Date/Time : 2024:03:11 16:27:04+01:00
Image Width : 1024
Image Height : 1536
SRGB Rendering : Perceptual
Exif Byte Order : Big-endian (Motorola, MM)
Image Description : 10.03.2024, Großbritannien, Windsor: Das undatierte, vom Kensington-Palast herausgegebene Handout-Foto zeigt Kate, Prinzessin von Wales, mit ihren Kindern, Prinz Louis, Prinz George und Prinzessin Charlotte, aufgenommen in Windsor, Anfang dieser Woche, vom Prinzen von Wales. Prinzessin Kate bedankte sich in einer Botschaft in den sozialen Medien für die anhaltende Unterstützung und wünschte den Menschen einen schönen Muttertag. Foto: Prince Of Wales/Kensington Palast/PA Media/dpa – ACHTUNG: Nur zur redaktionellen Verwendung bis zum 31.12.2024 und nur mit vollständiger Nennung des vorstehenden Credits. Das Foto darf nicht bearbeitet oder im Ausschnitt verändert werden. +++ dpa-Bildfunk +++
Artist : Prince Of Wales
Exif Version : 0232
Date/Time Original : 2024:03:10 02:34:23
Create Date : 2024:03:10 02:34:23
Source : Kensington Palast/PA Media
Urgency : 4
Transmission Reference : 911-004243
Instructions : UNITED KINGDOM OUT, IRELAND OUT, PICTURE DESK USE ONLY. NO SALES. HANDOUT
Supplemental Categories : Leute
Credit : dpa
Caption Writer : kde
Title : urn:newsml:dpa.com:20090101:240310-911-004243
Elvis ID : 9WexS6c3amM9b0m_iOBING
Keyword : Monarchie, Royals, Familie
Credit Line : dpa
Image Size : 1024×1536
Megapixels : 1.6
Situation: credible, should show their well being
Photographer: allegedly husband
Camera: unknown, cropped wide angle?
The overall look: A bit weird and plastic look in my eyes. Dimensions are wrong as her upper body seems too large for her legs. The right arm of Louis (and even George?) seem too long. Trying the posture of her in reality shows that it is unreal to get the embracing hands in this position.
Hands: The fingers of Charlotte’s left hand are larger than the fingers of her right hand. The index finger of Louis is missing which is difficult to reproduce in front of a mirror.
Teeth: Kate’s teeth look authentic when compared with other pictures of her, except for an unsharp band on the upper front teeth. Without having other images at hand, the teeth of the children look age-related (although Louis may be older than 5 on this picture).
Pattern: Floor looks good except left wall. There is a gap at the patten of right arm of Louis and lower left arm of Charlotte. While images of natural objects never have 100% identical patterns, such patterns are frequent with man-made objects — and difficult for AI to reproduce.
Sharpness: Floor mosaic: gets unsharp from tile 6 onwards – which is otherwise a perfect sharp area in the rest of the image.
Irregular: Kate’s right upper shoe border looks double. Zipper misaligns.
Sun/Shadow: sunshine on Kate’s left hand although there should be shadow under Charlotte’s arm. And well there background in the triangle under Charlotte’s arm is missing. The window mirror shows a tree that could cast more shadow on the scene.
General: Green leaves on the background trees in early March?
no clear results from splicing probability heatmap and ELA
just 4 hours later a new message that I still do not believe to be the whole truth https://twitter.com/KensingtonRoyal/status/1767135566645092616
Maybe that business should be left to professional photographers?
More comments at SPON by Matthias Kremp: Another possibility is a Google Pixel 8 that combines internally images which is however unlikely here. Kremp notices also the white paint at the step behind Louis.
SKY believes that The first save was made at 9.54pm on Friday night, with the second at 9.39am on Saturday morning.
The image was taken at Adelaide Cottage – the family’s home in Windsor – on a Canon 5D mark IV, which retails at £2,929.99 and used a Canon 50mm lens, which is priced at £1,629.99.
which contradicts the dpa exif data…
March 20, 2024
Guardian “Photo of Queen Elizabeth II and family was enhanced at source, agency says” and a famous photographer Pete Souza “lets not call it photoshopped”.
Doing now another image integrity study, I fear that we may already have the deep fake images in current scientific papers. Never spotted any in the wild which doesn’t mean that it does not exist…
Here are some T cells that I produced this morning.
The Journal of Allergy and Clinical Immunology (JACI) published numerous images that have been heavily modified. For Pubpeer examples see 1, 2, 3, 4, 5, 6, 7…
As I learned recently, this happened without the knowledge of the authors. As the duplications are always at the outer edges and originate from the same image, we assume that JACI publishing service inserted some cloned background” to insert”journal style” letter boxes.
Difficult to understand? Here is an example from the first reference.
…the duplications occurred when the publisher’s compositor vendor styled the figure panel label for print. The error did not impact the analysis, results, or conclusions of the article. The original figure appears below. The authors were not responsible for the error. Measures have been taken to prevent reoccurrence of this error for which the publisher accepts full responsibility and apologizes.
Although PDFs are the main formal output nowadays (no more “papers”) there is basically no standardization which meta data should be included in scientifc PDFs. It’s largely due to the software in the production office what’s in the PDF document — a largely underused resource.
This is at least my experience when working on an image duplication pipeline in scientific papers. But test it yourself ….
Blockiness in “busy” regions – may mimick duplications
Gibbs phenomenon – small highlighted area of large contrast
Alias, Moiré, frequent color/brightness change
Now let’s have a look at the compression artifacts in descending quality (BTW this also good exercise for the upcoming 125kB Content Upload Filter, I am therefore adding filesize also)
I am summarizing here some rules what consists of acceptable image use in scientific publications (Rossner 2004). Pioneers in the field have been the Journal of Cell Biology (J) but also Nature (N) and EMBO Press (E). Here are 10 rules:
Digital images submitted with a manuscript for review should be minimally processed (N).
Specific feature within an image may not be enhanced, obscured, moved, removed, or introduced (J, E). The use of touch-up tools, such as cloning and healing tools, or any feature that deliberately obscures manipulations, is unacceptable (N).
Dividing lines may not be added between juxtaposed images taken from different parts of the same gel or from different gels, fields, or exposures (J, E). If juxtaposing images is essential, the surrounding gel shown at least at the size of the band, borders should be clearly demarcated and described in the legend (N).
Images gathered at different times or from different locations should not be combined into a single image, unless it is stated that the resultant image is a product of time-averaged data or a time-lapse sequence.
Adjustments of brightness, contrast, or color balance that have been applied to the entire image may not enhance, erase, or misrepresent any information present in the original, including the background (J, E).
Images from the same object may not be repeated within the manuscript. Any reuse of images, including control images from earlier papers, should be explicitly stated and justified in the legend (J) as reuse would pretend any independent evidence. Reprobed Western blots should be clearly indicated.
Nonlinear adjustments (e.g., changes to gamma settings) must be disclosed in the figure legend (J). Processing such as changing brightness and contrast is appropriate only when it is applied equally across the entire image and is applied equally to controls. Contrast should not be adjusted so that data disappear (N).
Excessive manipulations, such as processing to emphasize one region in the image at the expense of others, is inappropriate (N).
Deposition of all RAW image files is encouraged (N). An option is figshare.com that is available since 2014.
There should be a better distinction of object and image, e.g. the re-use of the same object in multiple images should be banned. With any additional image in a paper, this is assumed to be additional and independent evidence of a given fact. Exception of that rule is a clear label (box, arrows, letters) as well as a legend why and from where it has been taken from (J).
These are general rules only. Specific rules exist for electrophoretic blots and microscopy (N). Additional points
There should be a reference to any work outside of the current paper, if the object and/or image has been used elsewhere.
There is a need for better captions that specify always experimental conditions, object, direction and resolution. Unzoomed controls and reason for the selection of certain object areas along with repeated experiments would be often helpful. The question is largely unsolved what constitutes a representative image?
There is a need of professional integrity offices at universities and research centers.
