I've got to give major props to Victoria's Secret. While they do photoshop their images, they take criticism well. When they were criticized for photoshopping out the shadows from one of
their bikini models, they quickly revised the image and left in the shadows. They
still did some edits, but it did not look as bad.
The reaction by Victoria's Secret is in stark contrast to Ralph Lauren, who grossly photoshops their models and attempts to
silence criticism by filing
DMCA takedown notices.
Form Fitting
Following the criticism by
Photoshop Disasters (PsD) and
myself, Victoria's Secret has
revised the image.
The original image was observed by PsD as missing a handbag. The new image contains the handbag.
Old: Handbag is missing.
| 
New: Handbag is present |
As well as adding back in the handbag, they addressed many of my discoveries including much of the modified skin. For example:
Old
| 
New |
A comparison between the old and new luminance gradient (LG) shows:
- Forehead: The old LG identifies a distorted forehead. The distortion is due to digital manipulation; they smoothed out her forehead and recolored it. The new image has a natural forehead. There is no odd ridge and the bumpiness looks like realistic noise.
- Elbow: The old image has a sharp curve in the rear arm's elbow. The curve lacks any bumps or noise. This is artificial. The new arm still has the curve (due to the light hitting it), but it is bumpy and realistic.
- Forearm: The cut on the old forearm went into the bicep. The new cut is at least straight (removing the shark-bite). LG also shows more details and realistic noise in the new image.
Of course, both images still have an artificial blur on the background and some recoloring on the skin, but the new image does look more realistic.
Or Does It?
Of course, most graphic artists cannot just release a picture. They must make changes. And when they rush their changes, they end up making mistakes. As computer forensic specialists, it is these mistakes that tell us what really happened.
The big question is: What did the original dress look like? Victoria's Secret permits users to change the dress colors. The new colors are still artificial, but much better. For example, here's what the red dresses look like. The new red is closer to realistic and the dress itself contains much more detail.
Old Red
| 
New Red |
Of all of the colors available, I think the original dress looked like the brown dress but was actually white. All of the other colors (not brown) were derived from the white version. Here's how I reached this conclusion:
New Brown
| 
New White |
Notice how the brown dress has much deeper fabric folds and realistic wrinkles than the white dress (and all of the other color dresses). These were the original wrinkles and were digitally removed for the other images. This even appears in the Principal Component Analysis (PCA). The brown image shows lots of JPEG artifacts, indicating multiple resaves while the white dress lacks these resave artifacts.
PCA of New Brown
| 
PCA of New White |
The PCA tells us what the dress looked like before the second save. However, the color of the original dress comes from a much more subtle clue. Look at the hair on the brown dress. Specifically, the curve of hair on her shoulder. There is a white chunk in the hair. On the brown dress, it looks like a very bright specular reflection. However, when seen on the white dress, it is obviously part of the white dress. When the artists selected the dress for recoloring, they forgot to select the small section of white in the hair. (In the old image, the artists did not make this mistake. But in the rush to create the new image, they screwed up.)
The Takeaway
Whether it is image analysis, keyboard profiling, or file recovery, there is one thing that I really enjoy about computer forensics: While the attacker (or artist) must cover all evidence in order to remain hidden, the forensic examiner only needs to find one clue.
In this case, the graphic artists at Victoria's Secret rushed the picture and left clues about what work was performed. They blurred the background and recolored the brown dress, then saved the image. Then they edited the dress to remove wrinkles and recolored every dress and resaved them again -- except for the brown one which was saved a second time with no additional edits.
From a truth-in-advertising viewpoint, this is misleading and potentially fraudulent. They show you one thing, but sell you something else. As a good example, consider the red dress. The old red was a different color than the new red. We don't know what the real red dress looks like since both old and new have been digitally colored and digitally enhanced.
From a "women's bodies" viewpoint, they are implying that there must be something wrong with you if the dress does not look that nice when you wear it. It's not you, it's not the dress. It's the fake picture of the dress.
I was watching an old Fred Astaire/Audrey Hepburn movie called, I think, "Funny Face". In it, Astaire plays a fashion magazine photographer. During one scene, a model is posing with a piece of sculpture. When the camera angle changes, you can see a line of clothes pins up her spine that hold the cloth of the dress tightly against her body in front.
It is not just dark skinned models. I've seen light skinned models recolored, too.
By the way, you can still see a little of her natural skin color in this picture: look at the dark shadow on her neck, and top of her ear. It isn't all shadow.
You can verify this yourself by downloading the new and old images and:
1. Scale the new image to 420x566. (99%x99% of the image, or 1% smaller)
2. Overlay it with the old image.
It also removed the shadow around the model's neck and some lines around her mouth.
I don't think the busts were digitally enhanced. This would have been too tedious to do without altering the fabric lines (most visible on the brown shirt photo), at least for an amateur editor in a hurry. She was most likely wearing a padded bra. This is a lingerie company after all.
The fact that we can either only make statements about edits that can't be verified (because we don't have the "before" photo), or because the edit is obvious (going by visual analysis) lets me thing this whole endeavor is nothing more that some tea-leaves reading (where you see what you want to see, and confirm what's obvious)...
It does show up in the LG. You really should read the first blog entry about this picture (http://www.hackerfactor.com/blog/index.php?/archives/322-Body-By-Victoria.html) since it identifies lots of manipulations -- before the picture was updated.
But I have problems figuring out the differences between the old and new picture, regarding the sharp elbow curve, and the "shark bite", in the LG analysis.
I'm just not seeing the difference in the LG as you've described them.
I've always wondered how Photoshopped pictures are determined beyond the naked eye. Your previous article on this image was great!
I would love to hear one of your talks in person!