I find that Bing seems to do a lot of very similar female faces, often with glaring white teeth and lips that are very full. The eyeballs also seem out of whack in many cases. Overall, the faces look like a mannequin or doll, rather than human.
Here are three photos, a before-and-after close-up and a collage of pictures showing 'Bing Face' and the same collage put through Krea to make each face much more natural.
I'm at the point where recognizing 'Bing Face' takes the enjoyment out of viewing AI photos, but running it through Krea brings more of the human aspect in, making a noticeable improvement in the realism of the image.
The first collage is with Krea, the second one showing 'Bing Face'.
Is it just me, or are others here noticing 'Bing Face'?
As with any model, it's going to be skewed by the training data and also the prompts you use. Bing and Dalle3 also reinterpret the prompt you use so the prompt that the model sees/uses to generate the image has already been altered. This may also add bias towards a particular subject.
If your prompts use a similar description of the subject, then it's likely you will get similar subjects out for similar variations of the prompt. As an example, rather than just using "Beautiful Woman" each time, you need to embellish the prompt with adjectives or other descriptors. It's not perfect as other elements of the prompt will also reinforce some bias towards a particular look too. i.e. changing the outfit/clothing will impact the model chosen, even the name.
messg said: As with any model, it's going to be skewed by the training data and also the prompts you use. Bing and Dalle3 also reinterpret the prompt you use so the prompt that the model sees/uses to generate the image has already been altered. This may also add bias towards a particular subject.
If your prompts use a similar description of the subject, then it's likely you will get similar subjects out for similar variations of the prompt. As an example, rather than just using "Beautiful Woman" each time, you need to embellish the prompt with adjectives or other descriptors. It's not perfect as other elements of the prompt will also reinforce some bias towards a particular look too. i.e. changing the outfit/clothing will impact the model chosen, even the name.
This. You need to be creative to get realistic facial expressions. If you always describe a woman as just 'surprised', Bing's always going to give you the 'Bing face' as you call it.
But if she's an 'annoyed librarian, startled, disappointed, cringing with jolly exasperation', you might get something approaching a realistic reaction. Please don't take this prompt and run with it, this is just an example of the sort thing you can experiment with. You don't even need the word 'woman' or 'female' in your prompt if you use the pronoun 'she'.
All Krea does is use a different model to remix the image you have uploaded. For the sort of thing I want to make, most of the time for me it isn't an improvement. Of the hundreds of images I've uploaded, I've used it precisely once. To me, the Krea faces are hyperreal, with too much detail, exaggerated highlights and shadows, too much value contrast that is too 'equal' over the whole image, which makes things like wrinkles and eyelids look wrong to me. The Bing images are slightly 'posterised', and sometimes the eyes are wrong, but in many ways the lack of detail actually aids realism.
This. You need to be creative to get realistic facial expressions. If you always describe a woman as just 'surprised', Bing's always going to give you the 'Bing face' as you call it.
I am not even using any expression descriptions. I also notice other people's Bing images often have the same face. I don't mean expression, I mean it looks like it was modelled after the same person.
thereald said: To me, the Krea faces are hyperreal, with too much detail, exaggerated highlights and shadows, too much value contrast that is too 'equal' over the whole image, which makes things like wrinkles and eyelids look wrong to me.
I agree it can go that way, but there are settings in Krea which let you scale such things back to a more realistic level, between the plastic 'Bing Face' and the over hyper real Krea ones. The middle ground looks quite natural in most cases.
I think that there is the cause of it. All of my images are made using Dall-e 3 (either through BIng or Chat GPT), and the faces don't look like the ones you've uploaded.
'Woman' through Bing's dataset is going to result in someone who looks like an American soap actress, because it's based on Bing's image data. And that's what your Bing Face is.
'English Woman', or '27 year old gf', or 'annoying welsh lady', or any other number of things Ive tried in prompts, is going to get you different results.
thereald said: 'Woman' through Bing's dataset is going to result in someone who looks like an American soap actress, because it's based on Bing's image data. And that's what your Bing Face is.
'English Woman', or '27 year old gf', or 'annoying welsh lady', or any other number of things Ive tried in prompts, is going to get you different results.
Although I've steered away from 'shocked', 'surprised', 'laughing', etc. (because Bing does the expression based on the scene I've described) I do describe the woman, with features such as nationality, age, etc. and it makes some improvement. The eyes are almost always little spirals, and teeth too white, etc. on many Bing images I've seen from others too, not just my own.
My point was just to say that Krea has a nice fix by re-creating the faces and adding the missing character and human element. But when I look at the AI images here, I can recognize 'Bing Face' before even reading that it was done with Bing.
Here are three more examples of 'Bing Face' and the Krea improved versions. These were Bing's standard surprised expression, and I admit I could be more creative with this, but I still say the Krea improvement is amazing.
thereald said: 'Woman' through Bing's dataset is going to result in someone who looks like an American soap actress, because it's based on Bing's image data. And that's what your Bing Face is.
'English Woman', or '27 year old gf', or 'annoying welsh lady', or any other number of things Ive tried in prompts, is going to get you different results.
Although I've steered away from 'shocked', 'surprised', 'laughing', etc. (because Bing does the expression based on the scene I've described) I do describe the woman, with features such as nationality, age, etc. and it makes some improvement. The eyes are almost always little spirals, and teeth too white, etc. on many Bing images I've seen from others too, not just my own.
My point was just to say that Krea has a nice fix by re-creating the faces and adding the missing character and human element. But when I look at the AI images here, I can recognize 'Bing Face' before even reading that it was done with Bing.
I think your missing the point we were making.. All models have an inherent bias based on the data set used. With Bing, various elements of the prompt will reinforce this bias. It's not just asking for nationality, age, expression that impacts it. Things like English Rose beauty, Motherly Cosmopolitan, Ethereal beauty. etc..
Additionally, even the outfit can make a significant difference. It's hard to explain precisely but think about each word of your prompt.. Lets say your prompt is 400 characters long. When you generate it will be given a random seed number and an image is generated. When you click again, you will get a different seed number and a different image. However, the words (Characters/tokens) remain the same so the variations of the output of the different seeds are still largely the same.
Now.. Start changing the prompt.. A small change will start impacting the output. We can't choose the seed in Bing but you can in SD, Krea etc. If you could, you would be able to see exactly how much weighting Bing places on each of the tokens as it interprets the prompt. Because Bing also reinterprets the prompt you input before passing it to the image generation you loose a bit more of the control. It may be that despite your minor change, the rest of the prompt still influences the reinterpretation sufficiently that you still get a similar out. You need to make enough changes or more significant changes to impact Bing's 2nd prompt.
Krea.. I've used it quite a bit both for up-scaling, image generation and other features. Basically though it's a custom SD implementation with all the complex stuff hidden under the hood. Thats not to say it's terrible though. When you load an image to the up-scaler it uses an img2img prompt (positive/negative) and depending on the AI weighting and other settings will regenerate using the image as a guideline. The key point here is that Krea's own SD model will have different base set images so of course will get different faces. That said, you will see the same repetition in the Krea model if you were using it at high AI weightings or trying to generate a txt2Img output.
Basically, both tools are doing different things with different models so you are going to see different faces between them. Tighten up your bing prompt and you will see more variation, use Krea for up-scaling if needed afterwards. I also like Krea's inpainting features but better can be achieved using SD locally. Krita has AI plugins that work really well for this but Krita is a whole different discussion.
Great information, Messg! Much appreciated. I see what you are saying about the fine points of prompts and their effect on the outcome.
I suppose my main comment could have focused on the blurry/spiralled eyeballs Bing seems to produce, and the correction for them available through Krea, in addition to the fake white teeth and general 'mannequin' look in many Bing faces.
When I look at other people's images on here, I often see this right away, and know they used Bing, simply due to the eyeballs and teeth and overall plastic look of the faces. But of course, not always, as some are guiding Bing using more precise prompting. Thanks again for sharing your wisdom!
4/17/24, 4:17am: This post won't bump the thread to the top.
The models in bing photos that contain more than one woman will usually look like twins.
It depends how busy Bing is whether you get a woman which looks semi real or just total doll. The Krea tool does do a good job but usually you are polishing the Bing 'turd', you will still end up with girls who look like twins just much better looking overall.
Bing does produce a generic Asian woman. I can tell you the difference between Chinese and Japanese, (Pinay, Thai, etc) features as I've been to Asia many times. But still the generic Asian Bing woman is someone who looks like no country in particular.
Sometimes (rarely) you will get a black woman. I avoid specifying "black woman" cos that will trigger Bing racism police and get you banned.
Specifying "latina woman" usually sucks, they are typically unattractive. "spanish woman" produces someone who might come from generic South American country, not Spain.
"Italian woman" is my favorite though these all have beautiful long black hair and tan skin tone.
Bobographer said: Great information, Messg! Much appreciated. I see what you are saying about the fine points of prompts and their effect on the outcome.
I suppose my main comment could have focused on the blurry/spiralled eyeballs Bing seems to produce, and the correction for them available through Krea, in addition to the fake white teeth and general 'mannequin' look in many Bing faces.
When I look at other people's images on here, I often see this right away, and know they used Bing, simply due to the eyeballs and teeth and overall plastic look of the faces. But of course, not always, as some are guiding Bing using more precise prompting. Thanks again for sharing your wisdom!
Absolutely.. When Krea upscales it also fine tunes those areas of the image. My current workflow outside of SD would be a combination of Bing then Krea to upscale and fine tune. Bing used to be far easier to get realistic looks but as the model got censored, the workarounds needed to generate the images contaminates the image and in a lot of cases it just doesn't look real any more. I tend to focus my images on hyper realistic and prompting them is way more difficult than it used to be. It takes me much longer to find something that satisfies both the look of the model and also the realistic mess. That said, it still can be done. I've been working on a wedding theme over the last couple of days, various messes, outfits etc that I can use later to fine tune SD.
It's the wide angle lens look that always jumps out to me first with a lot of bing images. I'm not a user of it so not sure if that's because the prompt includes info down to the foot level and its going wide angle to fit everything in or if that is its default look. Wide can look more cinematic though sometimes which can make the image look like a shot from a TV show or movie but it stands the images quite apart from most realworld wam producers who tend not to shoot things so wide.
MMasia said: It's the wide angle lens look that always jumps out to me first with a lot of bing images. I'm not a user of it so not sure if that's because the prompt includes info down to the foot level and its going wide angle to fit everything in or if that is its default look. Wide can look more cinematic though sometimes which can make the image look like a shot from a TV show or movie but it stands the images quite apart from most realworld wam producers who tend not to shoot things so wide.
well spotted, I am/was using "Wide shot" specifically in the prompt for most of these images. I am working through a series of angles and shots. (Tilt, Medium, closeup, side profile etc) with the same underlying prompt.
MMasia said: It's the wide angle lens look that always jumps out to me first with a lot of bing images. I'm not a user of it so not sure if that's because the prompt includes info down to the foot level
I think it depends on different things, e.g. if you have a prompt only including description for the upper body and less about the surrounding scenery it tends to make portrait and close-up shots. But you can also add specific commands like "full body view" or "portrait of..." to the prompt which will get you different looks.