Annealing Generation: A Fire-Inspired Framework for Image Synthesis Through Pixel Evolution, Cooling, and Extinction

Abstract

Image generation methods commonly construct images through iterative denoising, autoregressive prediction, or transformations of learned latent representations. This paper proposes Annealing Generation (AG), a conceptual framework that treats image synthesis as a process analogous to the ignition, evolution, cooling, and extinction of fire. Rather than considering generation exclusively as the progressive removal of noise, AG models image formation as the evolution of a potentially variable population of candidate pixel entities. During the ignition and high-activity stages, candidate pixels may be created, relocated, recolored, merged, split, or removed. As the process cools, a temperature-conditioned controller progressively regulates these operations, encouraging the emergence and stabilization of coherent image structures. During extinction, the system terminates active evolution and decodes the remaining stable configuration into a final image. The framework distinguishes the metaphor of annealing from its operational implementation by defining temperature-dependent transition policies, candidate-population dynamics, an image-quality objective, and a stability-based stopping criterion. A mathematical formulation and a proposed learning procedure are presented, together with an experimental protocol for comparing AG against conventional diffusion generation and other relevant baselines. The paper does not claim that annealing-based image generation, dynamic pixel populations, or stochastic image optimization are individually new. Instead, it identifies the integration of fire-inspired pixel evolution, temperature-controlled population dynamics, and explicit extinction-based termination as a research hypothesis requiring implementation and empirical validation. The proposed framework aims to investigate whether this alternative formulation can provide useful control over image formation, computational expenditure, and structural stability.