Research map / Image generation / Fast diffusion sampling

EasyInv: Toward Fast and Better DDIM Inversion

ICML 2025 · Image generation › Fast diffusion sampling · approach: Distillation for few-step generation · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes EasyInv, a training-free DDIM inversion variant that blends the current latent with the previous one at selected timesteps (Kalman-filter-inspired) to avoid iterative noise refinement.

Problem

DDIM inversion yields inaccurate noise estimates, degrading reconstruction; existing fixed-point/iterative refinements (ReNoise, Pan et al.) improve quality but cost multiple model passes per step and can fail on special inputs.

Approach

Instead of iteratively optimizing per-step noise, periodically aggregate the current inverted latent with the previous latent (z_t = eta*z_t + (1-eta)*z_{t-1}) at selected steps, interpreted as a simplified constant-gain Kalman filter, reinforcing the initial latent's influence at no extra model calls.

Key results

• On SD v1.4 with 2,298 COCO images, EasyInv reaches SSIM 0.646 (best) and PSNR 30.189 at 5s, versus ReNoise PSNR 31.025 at 16s and DDIM 29.717 at 5s.
• EasyInv matches DDIM inversion's runtime while being about 3x faster than ReNoise (16s) and Fixed-Point Iteration (14s).
• Adding EasyInv to DirectInv improves 5 of 7 metrics across PnP-Inversion editing tasks (e.g., structure distance 24.29 to 22.88 for PnP).

Measured effects reported in the paper

Surprising findings

• ReNoise produces black images on inputs with large white areas, and fixed-point iteration and DDIM also fail on such cases.
• The method's reported benefit is largest for weaker models; the iterative Pan et al. method works on SD-XL but fails on SD v1.4.

Limitations

• Over-denoising can produce overly smooth outputs, as in the peach example.
• Hyperparameters (eta=0.5, selected steps 0.05T to 0.25T) are chosen empirically; the Kalman gain is treated as constant rather than computed.
• Results on the PnP-Inversion table are mostly taken from prior work and the in-text comparisons are on a single model/setup.

Experimental setup

Datasets: COCO 2017 test and validation sets (2,298 images) PnP Inversion editing dataset (9 editing tasks)

Benchmarks: COCO 2017 subset reconstruction PnP Inversion editing benchmark

Metrics: LPIPS SSIM PSNR inference time structure distance MSE CLIP similarity

Baselines: DDIM Inversion ReNoise Fixed-Point Iteration Null-text inversion Negative-Prompt Inversion StyleDiffusion DirectInv

Builds on: DDIM ReNoise Fixed-point iteration inversion (Pan et al.) Null-text inversion DirectInv

Closest papers by meaning

• FreeInv: Free Lunch for Improving DDIM Inversion (NeurIPS 2025 · similarity 0.91)
• BELM: Bidirectional Explicit Linear Multi-step Sampler for Exact Inversion in Diffusion Models (NeurIPS 2024 · similarity 0.85)
• Adapt and Diffuse: Sample-adaptive Reconstruction via Latent Diffusion Models (ICML 2024 · similarity 0.85)
• Sparse Model Inversion: Efficient Inversion of Vision Transformers for Data-Free Applications (ICML 2024 · similarity 0.84)
• Exploring Fixed Point in Image Editing: Theoretical Support and Convergence Optimization (NeurIPS 2024 · similarity 0.84)
• Integrating Intermediate Layer Optimization and Projected Gradient Descent for Solving Inverse Problems with Diffusion Models (ICML 2025 · similarity 0.84)