They aren't getting increasingly yellow (I don't think the base model has been updated since the release of GPT-4o Image Generation), but the fact that they are always so yellow is bizarre and I am still shocked OpenAI shipped it knowing that effect exists, especially since it has the practical effect of instantly being able to clock it as an AI image generation.
Generally when training image encoders/decoders, the input images are normalized so some base commonality is possible (when playing around with Flux Kontext image-to-image I've noticed subtle adjustments in image temperature), but the fact that it's piss yellow is baffling. The autoregressive nature of the generation would not explain it either.
Generally when training image encoders/decoders, the input images are normalized so some base commonality is possible (when playing around with Flux Kontext image-to-image I've noticed subtle adjustments in image temperature), but the fact that it's piss yellow is baffling. The autoregressive nature of the generation would not explain it either.