In the standard paradigm of diffusion models, the "forward process"—the transformation of structured data into unstructured noise—is treated as a fixed physical law. Whether using a linear schedule as in the original DDPM or a cosine schedule, these models typically rely on a predefined, data-agnostic sequence of Gaussian noise additions.
However, a growing body of research suggests that this "one-size-fits-all" approach is a bottleneck. By treating the noise schedule and the forward trajectory as learnable parameters rather than fixed hyperparameters, researchers are finding they can achieve higher sample quality, faster inference, and better density estimation.
The Limitation of Fixed Destruction
The intuition behind standard diffusion is that any data distribution can be mapped to a simple Gaussian distribution if you add enough noise. But as Grigory Bartosh argues in Neural Flow Diffusion Models, these fixed, linear paths often force the reverse (generative) process to learn highly complex, curved trajectories.
Imagine trying to navigate from a mountain peak (data) to a specific point in a valley (noise). If the path down is dictated by a rigid, external rule, you might have to take a winding, inefficient route. If you can instead choose how to descend, you can create a straighter, simpler path that is much easier to retrace during generation.
Learning How to Noise
One major facet of this research direction is the Learnable Forward Process. Instead of adding noise blindly, models like Neural Diffusion Models and DiffEnc (by Anders Christensen) introduce time-dependent, non-linear transformations.
By allowing the model to "pre-process" or "simplify" the data before noise is even added, the subsequent denoising task becomes significantly easier. In the discrete domain, Forward-Learned Discrete Diffusion (FLDD) demonstrates that learning how to "corrupt" text or molecules allows for high-quality generation in as few as 10 steps, compared to the hundreds required by fixed-masking schedules.
Instance-Specific and On-the-Fly Adaptation
While some models learn a better schedule for an entire dataset, others, like the Time Prediction Diffusion Model (TPDM) by Zhiyang Chen and colleagues, adapt the schedule on-the-fly for every single sample.
The core insight here is that not all images are equally difficult to generate. A simple prompt (e.g., "a blue square") might require only a few denoising steps, while a complex scene (e.g., "a crowded cyberpunk city") requires many more. TPDM uses reinforcement learning to predict the optimal next noise level for a specific image, cutting sampling steps by 50% without sacrificing quality. Similarly, Carlos Esteves and Ameesh Makadia propose Spectrally-Guided Diffusion, which tailors the noise level to the unique power spectrum of each individual image.
Multivariate and Per-Pixel Control
Standard models apply the same amount of noise to every pixel or dimension at a given timestep. However, different features of a dataset often have varying levels of complexity or semantic importance.
- MuLAN (Multivariate Learned Adaptive Noise): Aaron Gokaslan and Volodymyr Kuleshov introduced MuLAN, which learns a per-pixel, data-dependent noise schedule. This allows the model to focus its capacity on the most difficult-to-model parts of an image, leading to state-of-the-art density estimation.
- Asynchronous Denoising: Zijing Hu and colleagues developed AsynDM, which allows background regions to clear up faster than the primary objects. This provides a stable "context" for the model, improving the alignment between the generated image and the text prompt.
Why It Works: The Geometry of Transport
The success of these methods—from BDDM in speech synthesis to Equivariant Neural Diffusion in molecule generation—stems from a shared principle: Optimal Transport.
By moving away from fixed SDEs and embracing learnable trajectories, these models are essentially solving a more flexible version of the Schrödinger Bridge problem. As explored in LDSB by Xingyu Qiu, finding the "local" optimal path between noise and data distributions minimizes the effort the neural network must expend to reverse the process. When the path is straighter and the noise levels are calibrated to the data's actual geometry, the model can take larger, more confident steps, leading to the massive speedups and fidelity gains seen across this research direction.