EasyInv Noise Refinement (BDIA)

• EasyInv Noise Refinement (BDIA) is a technique that refines noise estimation in diffusion models by symmetrically combining forward and backward integrations for exact inversion.
• It achieves round-trip consistency in image editing by using one neural network evaluation per step, avoiding the state-inconsistency seen in standard DDIM inversion.
• Empirical benchmarks demonstrate that BDIA improves FID scores in text-to-image and unconditional generation tasks, while generalizing effectively to higher-order ODE samplers.

EasyInv Noise Refinement—also known as Bi-Directional Integration Approximation (BDIA)—is a technique for refining noise estimation in diffusion models to achieve exact and consistent inversion, particularly in the context of diffusion-based generative samplers such as DDIM and related ODE solvers. BDIA addresses state-inconsistency in DDIM inversion for image editing, enabling round-trip consistency at negligible computational overhead by forming a time-symmetric, algebraically invertible update using both forward and backward integration approximations. This yields improved sample quality in both unconditional and conditional generation settings and generalizes naturally to higher-order ODE samplers (Zhang et al., 2023).

1. Definition and Objectives

BDIA refines the noise estimation in numerical diffusion inversion by replacing the traditional first-order one-way update found in DDIM sampling with a time-symmetric rule that combines forward and backward steps using the same neural network evaluation. The method computes the next state $z_{i-1}$ at time $t_i$ using the current estimate $z_i$, the preceding/following hidden state $z_{i+1}$, and a single model estimate of the noise $\hat{\epsilon}(z_i, i)$. The principal objectives are to:

• Enable exact round-trip (invertible) diffusion inversion for image editing.
• Avoid the state-inconsistency present in standard DDIM inversion.
• Achieve these goals with only one neural network evaluation per step, in contrast to alternatives like EDICT (which requires two calls per step) and null-text inversion (which requires iterative optimization) (Zhang et al., 2023).

2. Derivation of Core Update Rule

Let $a_i = \frac{\alpha_{i-1}}{\alpha_i}$ and $$b_i = \sigma_{i-1} - \sigma_i \frac{\alpha_{i-1}}{\alpha_i}$$ denote the DDIM update coefficients. The forward increment is defined as

$$\Delta_{\rm f}(i \rightarrow i-1) = z_{i-1} - z_i = a_i z_i + b_i \hat{\epsilon}(z_i, i) - z_i$$

The backward increment, reusing $\hat{\epsilon}(z_i, i)$, is

$$\Delta_{\rm b}(i \rightarrow i+1) = z_{i+1} - z_i = \alpha_{i+1} \left( \frac{z_i - \sigma_i \hat{\epsilon}(z_i, i)}{\alpha_i} \right) + \sigma_{i+1} \hat{\epsilon}(z_i, i) - z_i$$

The BDIA update is then formed by the symmetric combination

$$z_{i-1} = z_{i+1} - \Delta_{\rm b}(i \rightarrow i+1) + \Delta_{\rm f}(i \rightarrow i-1)$$

or explicitly,

$$z_{i-1} = z_{i+1} + \left(a_i - \frac{1}{a_{i+1}}\right) z_i + \left(b_i + \frac{b_{i+1}}{a_{i+1}}\right) \hat{\epsilon}(z_i, i)$$

This update is linear in $z_{i+1}, z_i, \hat{\epsilon}(z_i, i)$, making it directly invertible.

3. BDIA-DDIM Sampling Algorithm

The BDIA extension to DDIM is realized with only minor additions to the standard algorithm and no increase in neural network evaluation cost. The procedure is as follows:

• Precompute $a_i, b_i$ for the schedule $t_N > \ldots > t_0$.
• The first step computes $z_{N-1}$ via the standard DDIM update from $z_N$.
• For subsequent steps, for each timestep $i$ from $N-1$ down to $1$:
  • Obtain $\hat{\epsilon}(z_i, i)$.
  • Compute forward increment $\Delta_{\rm f}$.
  • Compute backward increment $\Delta_{\rm b}$.
  • Update by $$z_{i-1} = z_{i+1} - \Delta_{\rm b} + \Delta_{\rm f}$$.

This process requires only two vector additions and a few scalar multiplications beyond DDIM, with a single neural network evaluation per step.

4. Exact Inversion and Computational Considerations

Because the BDIA-DDIM update produces a linear combination of $z_{i+1}, z_i$, and $\hat{\epsilon}(z_i, i)$, it admits exact algebraic inversion: given $(z_{i-1}, z_i)$, $z_{i+1}$ is recovered by

$$z_{i+1} = z_{i-1} + \Delta_{\rm b}(i \rightarrow i+1) - \Delta_{\rm f}(i \rightarrow i-1)$$

This property enables reversible, consistent image editing and round-trip inversion, avoiding the state inconsistency observed in standard DDIM inversion. Compared to EDICT (which maintains an auxiliary chain and doubles the inference cost) and null-text inversion (which involves computationally expensive iterative optimization at every step), BDIA achieves consistency with only algebraic overhead and no increase in neural model calls (Zhang et al., 2023).

5. Generalization to Other ODE-Based Samplers

The core principle of BDIA—reusing the current step's model output to symmetrize integration across forward and backward time intervals—generalizes to higher-order ODE solvers. For EDM, which uses an improved Euler (Heun) step,

$\tilde{z}_{i-1} = z_i + (t_{i-1} - t_i) d_i$

$$z_{i-1} = z_i + (t_{i-1} - t_i) \frac{1}{2}(d_i + d_{i-1}')$$

where $d_i = d(z_i, t_i)$, BDIA-EDM forms $\hat{z}_i$ from $z_{i+1}, z_i$, and the gradients at adjacent time steps, before updating via the usual improved Euler rule. This again requires only one extra vector addition per step, with no increase in model evaluation count.

6. Empirical Performance and Applications

Benchmarking demonstrates that BDIA improves generative sample quality across diverse datasets and architectural backbones. Key results include:

• Text-to-Image (Stable Diffusion v2, 10 steps, COCO):
  • DDIM FID: 15.04
  • DPM–Solver FID: 16.06
  • BDIA-DDIM FID: 12.62
• Unconditional Generation (CIFAR10, CelebA; FID, DDIM vs. BDIA-DDIM):

Dataset	Steps = 10	20	40
CIFAR10	14.38 → 10.03	7.51 → 6.29	4.95 → 4.63
CelebA	13.41 → 10.86	9.45 → 8.86	6.93 → 6.50

• EDM vs. BDIA-EDM (56-sample FID):

Dataset	NFE	EDM	BDIA-EDM
CIFAR10 32²	35	1.85	1.79
FFHQ 64²	39	2.64	2.54
AFHQv2 64²	39	2.08	2.02
ImageNet 64²	39	2.51	2.38

• Round-trip Image Editing: BDIA-DDIM produces perceptually closer reconstructions and perfectly consistent editing trajectories, with only one neural model call per step—half the network evaluation cost of EDICT.

A user-tunable scalar $\gamma$ enables smooth trade-off between faithfulness and visual effect strength during editing.

7. Summary and Impact

BDIA (EasyInv Noise Refinement) is a universally applicable, computationally lightweight enhancement to DDIM and other diffusion ODE samplers. The method offers a time-symmetric, algebraically invertible update, exact diffusion inversion, more accurate sampling, and improved FID across multiple benchmarks. The cost is limited to a single additional vector addition per step, with no increase in neural network inference (Zhang et al., 2023). This suggests that BDIA represents a practical step forward for achieving efficient and consistent diffusion model inversion suitable for image editing and high-fidelity generative tasks.