On the Measurement of Realism in Synthetic Images. From Cornell Box Experiments to ImageNet and Inception Photorealism
Just as in the (older) field of computer graphics, a key goal in the development of generative image models has been the achievement of photorealism—i.e., the making of images that are perceived as indistinguishable from those produced with a camera lens. This paper locates current techniques for as...
保存先:
| 出版年: | URI:https://journals.openedition.org/transbordeur, |
|---|---|
| 第一著者: | |
| フォーマット: | Article ou chapitre numérique |
| 言語: | Anglais |
| 出版事項: |
Transbordeur
2025
|
| 主題: | |
| オンライン･アクセス: | Accès Université d'Orléans et IFPM Accès Université d'Orléans et IFPM |
| 要約: | Just as in the (older) field of computer graphics, a key goal in the development of generative image models has been the achievement of photorealism—i.e., the making of images that are perceived as indistinguishable from those produced with a camera lens. This paper locates current techniques for assessing photorealism in synthetic images within a longer history of efforts to evaluate realism in computer-generated content. More precisely, I compare two current standards for assessing photorealism in AI-generated images (the so-called FID metric and Inception score) with the assessment methods used in the Cornell box experiments—a compilation of evaluation techniques that set the standard for how realism was evaluated in computer-generated images in the 1980s and onwards. Exploring how ideas about realism and photorealism become translated into algorithmic models, and drawing on the works of Lev Manovich and Hannes Bajohr, I observe a shift from simulative and sequential enactments of photorealism to predictive and connectionist enactments of photorealism, starting in the period between 2014 and 2016. I also reveal how a specific AI training dataset (ImageNet) and neural network (Inception model) has played a central role in shaping the aesthetics of photorealism in generative image making since 2016. |
|---|