Forced examples of AI’s own image weaknesses, produces great abstract art.
AI image generation, while impressive, still struggles with certain consistent errors.
These are legion, including:
Human and Biological Form Errors – number & configuration of limbs, digits, teeth, eyes, hair and other anatomical feature distortions etc: (there be monsters, ever seen Carpenter’s movie ‘The Thing’, or the movie ‘Society’ ?)
Composition and Spatial Errors – Perspective, Geometry, Context, Symmetry, Form, Architecture, Proportion, Shape distortion etc: AI often just doesn’t understand the rational, 3D, Newtonian world we all inhabit
Detail, Lighting, Artifacts and Texture Errors – Shadows, Light sources, Reflections, Blurring, Compression artifacts, patterns etc: sometimes jarring mistakes can be seen in even the most photo-realistic creations
Other errors:- Complex scenes, text and symbols, simple counting, failure to understand unusual concepts, metaphor, impossible actions or events, lack of creativity etc: (*but see below)
Bias: As ever, there is the ‘dark passenger’ of AI generally: training data bias – including perpetuating stereotypes, misrepresentation of certain groups of people, misrepresenting facts, errors of chronology, association, causality, empirical evidence etc (AI Large Language Models work on enormous amounts of OUR existing data – we seed it)
*By way of experiment into AI’s self knowledge and self critique, we asked it (Gemini Imagen3) to example, in the form of an image, ALL of AI’s known image creation errors.
What it produced was quite revealing about how it interpreted our various prompts, and was certainly visually striking; a collection of fantastic, phantasmagorical creations akin to Dali or Hieronymus Bosch. (Although the genius of Dali and Bosch is not seriously challenged IMHO).
Examples below from Imagen3, using the prompt: “Create me an image which shows the major errors that AI often makes when generating images“.
Similar abstract images were also produced when asking for images based on a preceding text based answer for known errors. Contrary to our expectations, this long form method didn’t produce the more analytical or didactic renderings that we hoped for – but were nevertheless quite impressive, in their own way. The simple prompt above proves that ironically, sometimes less is more, in the instruction.
