I doubt you need an LLM for that, a diffusion model or perhaps even a deep CNN could probably do a passable job. You could train it by taking consecutive triplets of images [A, B, C] and providing [A, C] as inputs with B as the expected output.