Renderings like this seem very interesting to me. I feel like the photo looks realistic from afar but upon closer inspection something seems wrong about it. I was wondering if this feeling could be quantified into any specific feature or reason.
tom5079
Maybe there is an AI that can distinguish rendered images from real images?
aravmisra
I found this website: https://www.aiornot.com/
I know this isn't exactly what you both were referring to- AI vs simply rendered photos. I wonder if the process might be similar however? With SD, the process of breaking down each subsequent step into its liminal space probably hints at whether an image is AI-generated or not. Perhaps there's an equivalent here that models can use to detect?
rcorona
I'm not aware of any models used to distinguish images captured by cameras from rendered images, but to add to the discussion above, I think that training a model to do this could be complicated by the fact that there isn't a single canonical distribution for real images, since their properties will depend on the physical parameters of the camera used to capture each one.
For example, I'm reminded of this paper (https://arxiv.org/pdf/1505.05192.pdf) which presents a method for learning image features by having a model attempt to predict the relative positions of image patches. The intuition is that in order to learn how to predict relative positions, the model would need to learn about the appearance and structure of objects.
As a complication, the authors found that some cameras will refract light differently depending on which region of the lense it passes through, something which the model was leveraging to find a trivial solution to the relative position prediction problem (and hence was lowering the quality of learned representations).
Songbird94
This thread is super informative!
emily-xiao
Agree with zy5476 that the rendering feels ever so slightly off — particularly in terms of the way it captures stillness in parts that more realistically should be moving and the motion blur feels overdone. Also curious about details of the underlying model.
AbhiAlderman
I'm wondering what makes this photo appear so "off"? It seems almost realistic but certain elements make it seem AI-generated or rendered in some modeling software. I think it has something to do with the lighting on the flowers contrasting with the bees. The bees almost seem like cutouts pasted onto the page from a scene with a different light source. The motion blur on the wings also seems overdone, even though this is a hummingbird. Is it only the lighting that is impacting the realistic quality of this photo, or are there more elements affecting the quality?
CLV-Iclucia
I guess it's because the volumetric rendering of the flowers and furs are not handled well? Modelling these materials are always hard. I have seen many AI-generated faces and none of them can handle the SSS effect of skin correctly, as a result the faces look too smooth, as if they are made of plastics. I guess in this picture we have the same problem.
Renderings like this seem very interesting to me. I feel like the photo looks realistic from afar but upon closer inspection something seems wrong about it. I was wondering if this feeling could be quantified into any specific feature or reason.
Maybe there is an AI that can distinguish rendered images from real images?
I found this website: https://www.aiornot.com/
I know this isn't exactly what you both were referring to- AI vs simply rendered photos. I wonder if the process might be similar however? With SD, the process of breaking down each subsequent step into its liminal space probably hints at whether an image is AI-generated or not. Perhaps there's an equivalent here that models can use to detect?
I'm not aware of any models used to distinguish images captured by cameras from rendered images, but to add to the discussion above, I think that training a model to do this could be complicated by the fact that there isn't a single canonical distribution for real images, since their properties will depend on the physical parameters of the camera used to capture each one.
For example, I'm reminded of this paper (https://arxiv.org/pdf/1505.05192.pdf) which presents a method for learning image features by having a model attempt to predict the relative positions of image patches. The intuition is that in order to learn how to predict relative positions, the model would need to learn about the appearance and structure of objects.
As a complication, the authors found that some cameras will refract light differently depending on which region of the lense it passes through, something which the model was leveraging to find a trivial solution to the relative position prediction problem (and hence was lowering the quality of learned representations).
This thread is super informative!
Agree with zy5476 that the rendering feels ever so slightly off — particularly in terms of the way it captures stillness in parts that more realistically should be moving and the motion blur feels overdone. Also curious about details of the underlying model.
I'm wondering what makes this photo appear so "off"? It seems almost realistic but certain elements make it seem AI-generated or rendered in some modeling software. I think it has something to do with the lighting on the flowers contrasting with the bees. The bees almost seem like cutouts pasted onto the page from a scene with a different light source. The motion blur on the wings also seems overdone, even though this is a hummingbird. Is it only the lighting that is impacting the realistic quality of this photo, or are there more elements affecting the quality?
I guess it's because the volumetric rendering of the flowers and furs are not handled well? Modelling these materials are always hard. I have seen many AI-generated faces and none of them can handle the SSS effect of skin correctly, as a result the faces look too smooth, as if they are made of plastics. I guess in this picture we have the same problem.