Standard diffusion models are fundamentally asymmetric. To turn data into noise, they rely on a fixed, hand-crafted forward process—typically a linear or cosine schedule of Gaussian perturbations. Then, they task a highly expressive neural network with learning the reverse process to undo this damage.
This "blind destroyer" approach forces the generative network to adapt to an arbitrary trajectory. In high-dimensional spaces, a rigid, predefined noising path often forces the reverse process to navigate highly curved trajectories, passing through regions of low data density or creating sharp "velocity spikes" in the score function. To prevent numerical drift along these winding paths, standard models require hundreds of discretization steps during inference.
A rapidly growing body of research asks a simple question: Why not make the forward process smart too? By parameterizing the forward noising schedule, diffusion coefficients, or the entire destruction path as a learnable component, models can automatically discover optimal trajectories. This shift drastically reduces the number of sampling steps required for high-quality generation and yields tighter variational bounds for density estimation.
The Core Problem: Path Variance and Discretization Error
To understand why learnable forward processes work, we must look at what happens when a trajectory is suboptimal. When we diffuse data using a fixed heuristic, we are implicitly choosing an arbitrary bridge between the data distribution and pure noise.
If this bridge does not align with the intrinsic manifold of the data, the score function (the gradient of the log-density) changes rapidly and unpredictably across time steps. In A Minimum Variance Path Principle for Accurate and Stable Score-Based Density Ratio Estimation, Wei Chen et al. identify that these sharp transitions introduce a high "path variance" into the training objective. When the path has high variance, numerical solvers experience severe discretization errors unless they take tiny, computationally expensive steps.
By treating the interpolation path as a learnable component optimized for the specific dataset, the model can discover smooth, straight trajectories that bridge distributions with minimal overlap. Minimizing this path variance stabilizes training and allows the reverse process to be solved accurately in a fraction of the steps.
Parameterizing Destruction: From Schedules to Neural Flows
Researchers have tackled the challenge of learning the forward process at varying levels of flexibility, ranging from scalar schedules to full neural transformations.
1. Learnable Noise Schedules
The most straightforward approach keeps the Gaussian noise structure but learns the mapping of time to noise intensity. In On Density Estimation with Diffusion Models, Diederik P Kingma et al. parameterized the noise schedule using a monotonic neural network. This allowed the model to jointly optimize the path of data degradation alongside the denoising network, significantly improving likelihood-based modeling.
Similarly, Ishaan Gulrajani et al. demonstrated in Likelihood-Based Diffusion Language Models that optimizing the noise schedule via gradient descent minimizes loss variance, enabling diffusion models to achieve competitive likelihoods on text benchmarks. To eliminate manual hyperparameter tuning entirely, Gabriel Della Maggiora et al. introduced Conditional Variational Diffusion Models, which treats the variance schedule as a learnable parameter optimized via a variational objective, matching or exceeding the performance of manually tuned schedules in complex inverse problems.
2. Trainable Diffusion Coefficients
Moving beyond a global scalar schedule, models can learn dimension-wise or spatially varying noise levels. In Rethinking Losses for Diffusion Bridge Samplers, Sebastian Sanokowski et al. optimized diffusion coefficients using a reverse KL loss with a log-derivative estimator. This allowed the model to adaptively balance exploration and precision across different dimensions during training.
3. Fully Learnable Neural Transformations
The most flexible paradigm replaces fixed Gaussian noise entirely with data-dependent neural mappings. In Neural Flow Diffusion Models: Learnable Forward Process for Improved Diffusion Modelling, Grigory Bartosh et al. parameterized the forward marginal distribution using an invertible neural mapping. By allowing the forward process to adapt directly to the data structure, they achieved a tighter variational bound on the data distribution.
For discrete data, the same principle applies: Grigory Bartosh et al. introduced Forward-Learned Discrete Diffusion: Learning how to noise to denoise faster, utilizing a neural network to determine the noise distribution at each step conditioned on the entire input, making the subsequent denoising task significantly easier for the reverse model.
Adapting to Local Geometry
Why does a data-dependent forward process perform so much better? The intuition lies in how data is distributed in high-dimensional spaces. Real-world data typically concentrates on low-dimensional, highly curved manifolds. Fixed noise processes blast outward uniformly, ignoring this structure.
When the forward process is learnable, the model can adapt the intensity and direction of noise to the local geometry of the data distribution. As shown by Liam Hodgkinson et al. in Stochastic continuous normalizing flows: training SDEs as ODEs, a trainable diffusion coefficient acts as a form of adaptive regularization. The model can choose to inject less noise along directions of high data complexity to preserve sharp features, or smooth out complex densities where needed to improve sampling efficiency.
This behavior is explicitly leveraged in Adaptive Destruction Processes for Diffusion Samplers by Nikita Morozov et al., where parameterizing the forward destruction process allows the sampler to navigate complex energy landscapes with narrow, high-density regions—modes that fixed-variance processes frequently collapse or miss entirely.
The Payoff: Ultra-Fast Inference
The ultimate practical benefit of a learnable forward process is a massive reduction in the number of sampling steps required during inference. Because the forward path is optimized to be as straight and numerically stable as possible, the reverse process can take giant steps without drifting into instability.
Fixed Forward Process:
Data [Complex Manifold] ----(Winding, High-Variance Path)----> Noise
Requires ~1000 reverse steps to navigate the curves safely.
Learnable Forward Process:
Data [Complex Manifold] ====(Optimized, Straightened Path)====> Noise
Requires 3-10 reverse steps because the trajectory minimizes variance.
This acceleration is highly visible in speech and audio synthesis, where generation speed is critical. In BDDM: Bilateral Denoising Diffusion Models for Fast and High-Quality Speech Synthesis, Max W. Y. Lam et al. introduced a dedicated schedule network to parameterize both the forward and reverse processes. By finding a more efficient path from noise to data, BDDM achieved high-quality speech generation in as few as 3 to 7 steps.
Similarly, Rongjie Huang et al. developed FastDiff: A Fast Conditional Diffusion Model for High-Quality Speech Synthesis, which uses an auxiliary lightweight network to determine an optimized, condensed noise schedule for inference, reducing the required denoising iterations to just four steps.
By treating the destruction of data not as a static hyperparameter, but as a core part of the optimization problem, learnable forward processes bridge the gap between the flexibility of diffusion models and the efficiency of continuous normalizing flows.
Key Foundational Work
BDDM: Bilateral Denoising Diffusion Models for Fast and High-Quality Speech Synthesis (2022) — contributed 2 concept elements to this direction.
Stochastic continuous normalizing flows: training SDEs as ODEs (2021) — contributed 2 concept elements to this direction.
Learning Randomly Perturbed Structured Predictors for Direct Loss Minimization (2021) — contributed 2 concept elements to this direction.