Technical Summary: SynBoost: A Synergistic Framework for Fast Sampling of Diffusion Models

1. Problem Statement

Diffusion Probabilistic Models (DPMs) have achieved state-of-the-art results in visual generation tasks (image synthesis, editing, video, etc.). However, their practical deployment is hindered by slow inference speeds, as generating high-quality samples typically requires a large number of function evaluations (NFEs) due to the iterative sampling mechanism.

Existing acceleration strategies generally fall into two categories:

1. Training-based distillation: These methods reduce steps but require complex, model-specific distillation procedures.
2. Training-free fast samplers: These methods (e.g., DDIM, DPM-Solver, UniPC) leverage high-order solvers to minimize discretization error arising from approximating continuous integrals with large step sizes.

The authors identify a critical limitation in current paradigms: while high-order solvers effectively reduce discretization error, they largely ignore the approximation error caused by the neural network's imprecise estimation of the ground-truth score function. This oversight suggests that current optimization may have reached a plateau, leaving room for further acceleration by addressing the neglected error component.

2. Methodology

A. Error Decomposition and Dual-Error Disentanglement

The core theoretical contribution of the paper is the re-examination of the total sampling error. The authors decompose the error into two distinct components:

1. Discretization Error ($E_{dis}$): Arises from approximating the continuous exponential integrator with discrete steps.
2. Approximation Error ($E_{app}$): Arises from the neural network's inability to perfectly estimate the true score function (vector field).

To analyze these components, the authors propose a dual-error disentanglement strategy:

• They construct three transition processes over the same time interval $[t_{i-1}, t_i]$:
  • Exact: Derived from the pristine data distribution (error-free).
  • Approximation-only: Generated with an extremely fine-grained step size (e.g., NFE=1000) to minimize discretization error, isolating the approximation error.
  • Total Error: Generated with a coarse step size (e.g., NFE=10), containing both errors.
• By comparing these distributions using Mean Squared Error (MSE), they empirically demonstrate that:
  • Approximation error is of comparable magnitude to discretization error and often dominates at most timesteps.
  • Key Insight: Approximation error monotonically decreases as the timestep $t$ increases (moving from noise to data). This implies that predictions made at larger timesteps (closer to the initial noise) are more accurate regarding the score function estimation than those at smaller timesteps.

B. The SynBoost Framework

Based on the insight that approximation error decreases with increasing $t$, the authors propose SynBoost, a unified, training-free acceleration framework. It addresses both error sources simultaneously:

1. Mitigating Discretization Error: SynBoost seamlessly incorporates existing high-order ODE solver techniques (e.g., Taylor expansions used in DDIM, DPM-Solver, UniPC).
2. Mitigating Approximation Error: The authors introduce a mixing strategy that partially substitutes the current noise estimation with a more accurate prediction from a preceding, larger timestep $\tau$.

The modified noise estimation $\epsilon^{new}_\theta$ is formulated as:
$$\epsilon^{new}_\theta(x_{t_i}, t_i) = (1 + c)\epsilon_\theta(x_{t_i}, t_i) - c\epsilon_\theta(x_\tau, \tau)$$
Where:

• $c$ is a mixing coefficient that monotonically increases as the current step $t_i$ decreases (reflecting the need for more correction when approximation error is higher).
• $\tau$ is a larger timestep ($\tau > t_i$). In practice, the authors often use the initial noise $x_T$ (where $\tau = T$) as the source of the more accurate prediction, which is cached and does not incur extra NFEs.

This strategy is designed to be plug-and-play, compatible with various solvers (DDIM, DPM-Solver, DPM-Solver++, UniPC) and prediction modes (noise prediction and data prediction).

3. Key Contributions

• Error Disentanglement: The paper identifies and empirically isolates "approximation error" as a distinct and significant component of total sampling error, previously overlooked by solvers focusing solely on discretization.
• Monotonicity Discovery: The authors empirically establish that approximation error decreases monotonically as the timestep increases, providing a theoretical basis for using past (larger timestep) predictions to correct current estimates.
• SynBoost Framework: A training-free, unified framework that reduces total sampling error by blending current estimates with more accurate historical predictions.
• Broad Compatibility: The method integrates with existing high-order solvers without requiring model retraining or additional function evaluations per step.

4. Experimental Results

The authors validate SynBoost through extensive experiments across unconditional and conditional generation tasks, covering both pixel-space (ImageNet, LSUN) and latent-space (Stable Diffusion) models.

• Performance Metrics: Evaluated using FID (Fréchet Inception Distance) and HPD v2 (Human Preference Score).
• Key Findings:
  • Few-Step Regime: SynBoost significantly outperforms baseline samplers (DDIM, DPM-Solver, DPM-Solver++, UniPC) when NFEs are extremely limited (e.g., 5–10 steps).
  • Quality Improvement: Qualitative results show improved structural details, color contrast, and reduced artifacts compared to base solvers.
  • Compatibility: The method boosts performance across different orders (1st to 3rd order) and prediction modes. For instance, applying SynBoost to a 1st-order DDIM can yield results comparable to a 2nd-order DPM-Solver++.
  • Robustness: The framework remains effective under high classifier guidance scales (where samplers often become unstable) and on Transformer-based architectures (DiT).
  • Ablation: Sensitivity analysis confirms that a linearly decreasing strategy for the mixing coefficient $c$ and using the initial timestep $\tau=T$ yield optimal results.

5. Significance and Claims

The paper claims that SynBoost advances sampling efficiency beyond current paradigms by addressing the "under-explored" approximation error. The authors position the work as a significant step toward practical deployment of diffusion models, particularly in scenarios requiring very few sampling steps.

The authors maintain a modest stance regarding the limitations:

• SynBoost is a training-free method and currently lags behind training-based one-step generation methods (e.g., consistency models) in terms of ultimate speed and quality.
• The strategy is empirically driven and analyzed comprehensively, though the paper notes it does not pursue strict theoretical rigor in the derivation of the mixing strategy, relying instead on empirical evidence and mathematical derivation from exact DDIM steps.
• The work suggests that closing the gap between training-free and training-based methods requires future efforts.

In summary, the paper argues that by recognizing and correcting the approximation error inherent in neural network estimations, it is possible to substantially enhance the speed and quality of diffusion model inference without retraining the underlying models.