Jumping the Landing Phase: Noise Variance Matching Enables Accurate Few-Step Inversion

Among inversion approaches in diffusion models, DDIM in-version is widely used for reconstructing latent trajectories, offering adeterministic and training-free alternative to iterative methods. How-ever, we observe that it often suffers from reconstruction errors due to afundamental asymmetry in the inversion process: the noise is predictedfrom zt−1 rather than zt , leading to a consistent underestimation of noisemagnitude. This bias causes a partial denoising effect, where the latentstate is insufficiently perturbed, degrading reconstruction quality. Iter-ative methods such as ReNoise mitigate this issue by repeatedly refin-ing the predicted noise. We find that during the initial iterations, thepredicted variance rapidly increases toward its correct scale—an earlystage we term the landing phase, which accounts for most of ReNoise’scomputational cost. Based on this insight, we propose Noise VarianceMatching (NVM), a simple and efficient strategy that explicitly alignsthe predicted noise variance with that of a forward reference sample,bypassing the landing phase. NVM preserves the deterministic and skip-step inference property of DDIM while maintaining high reconstructionfidelity under few-step settings. Experiments across various datasets anddiffusion backbones demonstrate that NVM achieves competitive or su-perior reconstruction quality with far fewer UNet calls, offering a prac-tical and scalable alternative to iterative refinement-based methods.