Go to ICML 2025 Conference homepageEasyInv: Toward Fast and Better DDIM Inversion
This paper introduces EasyInv, an easy yet novel approach that significantly advances the field of DDIM Inversion by addressing the inherent inefficiencies and performance limitations of traditional iterative optimization methods. At the core of our EasyInv is a refined strategy for approximating inversion noise, which is pivotal for enhancing the accuracy and reliability of the inversion process. By prioritizing the initial latent state, which encapsulates rich information about the original images, EasyInv steers clear of the iterative refinement of noise items. Instead, we introduce a methodical aggregation of the latent state from the preceding time step with the current state, effectively increasing the influence of the initial latent state and mitigating the impact of noise. We illustrate that EasyInv is capable of delivering results that are either on par with or exceed those of the conventional DDIM Inversion approach, especially under conditions where the model's precision is limited or computational resources are scarce. Concurrently, our EasyInv offers an approximate threefold enhancement regarding inference efficiency over off-the-shelf iterative optimization techniques. It can be easily combined with most existing inversion methods by only four lines of code. See code at https://github.com/potato-kitty/EasyInv.
What is diffusion inversion? Diffusion inversion asks: “Given a pretrained diffusion model and a real image, can we recover the exact noise that produced that image?” If we succeed, feeding that noise back into the model reproduces the original photo. In other words, inversion is simply the reverse of the usual denoising process.
Why is this hard? A good inversion method must satisfy a “fixed‑point” requirement: at each diffusion timestep, the latent representation (the model’s compact encoding of the image) should be unchanged whether you invert then denoise or vice versa. Most existing algorithms enforce this by running multiple iterative refinements at every timestep—often three or more passes—which makes inversion several times slower than a single forward pass.
Our solution: EasyInv Inspired by the Kalman filter’s elegant way of blending new measurements with past estimates, we reuse the noise prediction from the previous timestep to inform the current inversion step. Crucially, we do this without any additional network calls or gradient computations or any other kinds of calculation. By slightly “dampening” each step’s noise update, we keep cumulative errors in check. A small correction early on compounds into a big boost in final accuracy—yet costs essentially zero extra time.
Why read the full paper? This simple explanation aims to provide a brief view of our paper and omits the mathematical derivation, like how Kalman Filter degenerate to our approach. For the complete information, please consult the paper itself.