BDIA: Exact Inversion in Diffusion Models

• BDIA is a technique that combines forward and backward ODE integration approximations to achieve exact inversion in diffusion models.
• It improves image sampling fidelity by reducing FID values (e.g., from 15.04 to 12.62 on COCO) while preserving computational efficiency.
• The method generalizes to alternative ODE solvers, enabling precise image editing and sub-pixel round-trip reconstruction.

Bi-Directional Integration Approximation (BDIA) is a technique introduced to address the inversion inconsistency present in Deterministic Diffusion Implicit Model (DDIM) sampling. It enables exact diffusion inversion with negligible computational overhead. BDIA constructs updates using both forward and backward ODE integration approximations, resulting in a closed-form, linear update for each latent state. The method allows precise recovery of forward latents from inversion queries and yields significantly improved sampling and editing fidelity compared to existing approaches, without incurring their computational costs. BDIA also generalizes to other ODE-based diffusion solvers, surpassing baseline performance in several experimental benchmarks (Zhang et al., 2023).

1. Context and Motivation

The DDIM framework permits fast, non-stochastic sampling in diffusion models by discretizing an ODE governing the latent process. In practice, DDIM trajectories are not exactly invertible due to their first-order nature, leading to significant mismatch—“inversion inconsistency”—between the original latent states $\{z_i\}$ and the states recovered via inversion $\{\hat{z}_i\}$. This discrepancy becomes particularly problematic for image-editing workflows, where drift from the initial content undermines precise reconstruction.

Previous remedies include Null-text inversion (Mokady et al.), which iteratively optimizes text embeddings through multiple gradient steps, and EDICT (Wallace et al.), which achieves exact inversion by introducing auxiliary latents and doubling network function evaluations per step. Both methods impose substantial computational overhead.

BDIA addresses this limitation by formulating a bi-directional integration procedure that retains the low per-step cost of DDIM but achieves exact inversion, explicitly correcting for first-order update errors while preserving the original sampling efficiency (Zhang et al., 2023).

2. Mathematical Formulation and Update Structure

Let $\alpha_i$ and $\sigma_i$ denote the noise schedule. The forward DDIM update is specified as

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

with $a_i = \alpha_{i-1}/\alpha_i$ and $$b_i = \sigma_{i-1} - \sigma_i (\alpha_{i-1}/\alpha_i)$$. This is a first-order ODE approximation for integrating over $[t_i, t_{i-1}]$.

In BDIA, the integration is enhanced by combining a forward step on $z_i$ and a backward step—originating from $z_{i+1}$ towards $z_i$. Introducing a parameter $\gamma \in [0,1]$, the BDIA-DDIM update is formulated as

$$z_{i-1} = \gamma (z_{i+1} - z_i) - \gamma \Delta(t_i \to t_{i+1} | z_i) + [a_i z_i + b_i \hat{\epsilon}(z_i, i)]$$

where $\Delta(t_i \to t_{i+1} | z_i)$ is the backward DDIM increment from $z_i$ to $z_{i+1}$. The special case $\gamma=1$ yields a time-symmetric form. Unwrapping this recursion, each $z_{i-1}$ is a linear combination of $z_{i+1}$, $z_i$, and $\hat{\epsilon}(z_i, i)$, thus affording explicit algebraic inversion for any desired direction.

3. Algorithmic Implementation

The BDIA-DDIM procedure proceeds as follows: For each discrete sampling step $i$ (from $N$ down to $0$),

• Compute the network-predicted noise $\hat{\epsilon}_i$,
• If $i < N$, evaluate the backward increment $$\Delta_\text{back} = \alpha_{i+1}((z_i - \sigma_i \hat{\epsilon}_i)/\alpha_i) + \sigma_{i+1} \hat{\epsilon}_i - z_i$$; else set $\Delta_\text{back} = 0$,
• Compute the forward increment $$\Delta_\text{forw} = a_i z_i + b_i \hat{\epsilon}_i - z_i$$,
• Update $z_{i-1}$ using the BDIA formula above.

The inversion step, given $(z_i, z_{i-1})$, exploits the linearity of the update: solving for $z_{i+1}$ in closed-form, without approximation, thereby achieving exact round-trip reconstruction up to floating-point error (Zhang et al., 2023).

4. Extension to Alternate ODE Solvers

BDIA generalizes beyond DDIM. For instance, Karras et al.’s EDM sampler employs an improved-Euler scheme:

$\tilde{z}_{i-1} = z_i + h_i d_i$

$$z_{i-1} = z_i + h_i (\frac{1}{2} d_i + \frac{1}{2} d'_{i-1})$$

where $d_i = d(z_i, t_i)$ and $d'_{i-1} = d(\tilde{z}_{i-1}, t_{i-1})$. In BDIA-EDM, the base point for the prior interval is refined to

$$\bar{z}_i = z_{i+1} + (1 - \gamma)(z_i - z_{i+1}) + \gamma (t_i - t_{i+1}) \cdot \frac{1}{2}(d_{i+1} + d_i)$$

with the improved-Euler update executed at $\bar{z}_i$ in lieu of $z_i$. This adjustment yields quantifiable performance improvements as measured by FID with no increase in neural network evaluations.

5. Experimental Evaluation

Empirical results validate BDIA’s improvements in both image generation and inversion fidelity:

Method	FID (COCO, StableDiffusion v2, 10 steps)
DDIM	15.04
DPM-Solver	16.06
BDIA-DDIM (γ=0.5)	12.62

On unconditional sampling benchmarks for CIFAR-10 and CelebA at $N=$10, 20, 40 steps, BDIA-DDIM outperforms standard DDIM consistently. For instance, on CIFAR-10 with 10 steps, FID improves from 14.38 (DDIM) to 10.03 (BDIA-DDIM). BDIA-EDM also yields consistent FID gains (0.1–0.2) across FFHQ, AFHQv2, and ImageNet64 at 39 steps.

Round-trip inversion using BDIA-DDIM achieves sub-pixel reconstruction error (MSE $\lesssim 10^{-5}$ at 40 steps). In guided image editing tasks, including text and ControlNet-based edits, BDIA-DDIM produces results that are perceptually closer to the original while halving the computational cost of EDICT and preserving its exact inversion property (Zhang et al., 2023).

6. Strengths, Limitations, and Prospects

Key strengths of BDIA include:

• Exact inversion requiring no iterative solves or auxiliary latents, maintaining a single $\hat{\epsilon}$ evaluation per step,
• Improved forward-sampling accuracy via backward correction, consistently yielding lower FID,
• Time-symmetric form at $\gamma=1$, assuring formulaic consistency between forward and backward propagation.

Limitations involve:

• A small additional arithmetic step per integration interval,
• The necessity to select $\gamma$, which mediates the trade-off between reconstruction fidelity and editing strength with no analytically optimal value.

Potential extensions encompass higher-order BDIA solvers making use of more temporal neighbors, adaptive per-timestep $\gamma$ scheduling to optimize output-target discrepancies, and application to other ODE-driven samplers such as DPM-Solver++, PLMS, and stochastic SDE solvers.

7. Significance and Research Directions

BDIA furnishes a principled, computationally efficient approach to exact inversion and improved sampling in diffusion models. It obviates reliance on costly iterative alignment or auxiliary latent strategies, thereby broadening the accessibility of high-fidelity inversion-driven editing and sampling workflows. This suggests further research may yield progressively more expressive and efficient solvers via higher-order integration schemes or by extending BDIA’s core concepts to alternative model classes and dynamic processes (Zhang et al., 2023).