Seeds and Reproducibility in AI Image Generation

Diffusion models start from pure random noise and refine it step by step into an image. That noise is not truly random. It comes from a pseudorandom number generator, which produces the same sequence of numbers whenever it starts from the same seed. For Stable Diffusion 1.5 at 512 by 512 pixels, the seed fills a latent tensor of 4 channels by 64 by 64 values, and everything the model draws grows out of that pattern. Large shapes, such as where a figure stands or where the horizon falls, are mostly decided in the early steps, so the seed strongly affects composition. Reproducing an image takes more than the seed. You also need the same model weights, prompt, negative prompt, resolution, sampler, step count and guidance scale, plus any extras such as LoRAs or ControlNet inputs. Changing the resolution changes the shape of the noise tensor, so the same seed then gives an unrelated image. Many tools give each image in a batch the next seed number, so the fourth image in a batch is often the starting seed plus three. Exact reproduction can still fail. The CPU and GPU random generators in libraries such as PyTorch produce different sequences from the same seed. Some GPU operations are nondeterministic, half-precision math rounds differently on different hardware, and memory-saving attention code can change tiny values that grow over many steps. Ancestral samplers add fresh noise at every step, which creates more chances for results to drift apart. A different model, or a new version of the same model, turns identical noise into a completely different picture. There are no universally 'good seeds'. A seed that works well for one prompt and model has no special quality anywhere else. Variation features, which blend in a second 'variation seed', let you explore nearby images while keeping the overall layout.