Zero Terminal SNR for Diffusion Models

32 concepts18 papers2 proposals

The core premise of a generative diffusion model is elegant: take pure, unstructured Gaussian noise and iteratively guide it toward a clean data distribution. At inference, this process begins at the final timestep ($t=T$), where the latent variable is assumed to be completely uninformative.

However, researchers have uncovered a fundamental mathematical discrepancy in popular diffusion implementations. In standard noise schedules (such as those originally used in Stable Diffusion), the process never actually reaches a Signal-to-Noise Ratio (SNR) of zero at the final training timestep. This subtle flaw introduces a "signal leak" that severely limits the model's expressive capabilities across images, video, and audio. Enforcing a zero terminal SNR fixes this blind spot, unlocking full dynamic range and eliminating persistent background artifacts.

The Intuition of Signal Leak

Why does a non-zero SNR at the final timestep matter? As Liu et al. and Everaert et al. demonstrated, if a noise schedule fails to completely destroy the image signal at $t=T$, a faint residual trace of the training data remains.

During training, the model learns to rely on this residual signal as a crutch. Because it is never exposed to a truly blank slate, it never learns how to establish global image properties—such as overall luminance, contrast, or dominant color—from scratch. Instead, when forced to make a prediction near $t=T$, the model defaults to the statistical mean of the entire training set.

Standard Schedule:  [Clean Image] ---> [Heavy Noise + Leaked Signal] 
                                                      |
                                                      v
                                        Model learns to rely on leak,
                                        defaults to mid-tone average.

Zero-SNR Schedule:  [Clean Image] ---> [Pure Gaussian Noise (SNR = 0)]
                                                      |
                                                      v
                                        Model forced to synthesize global
                                        structure & luminance from scratch.

At inference, we feed the model pure Gaussian noise. Because of the training discrepancy, the model interprets this empty starting point as a "leaked signal" representing a perfectly neutral, medium-gray image. This creates a severe color averaging effect or mid-tone bias, as detailed by He et al.. The model becomes mathematically incapable of generating extreme values, such as pure black or pure white scenes, resulting in a compressed dynamic range (Schusterbauer et al.).

Manifestations Across Modalities

The consequences of non-zero terminal SNR extend far beyond washed-out images, impacting any modality trained via diffusion:

  • Images and Sketches: High-contrast tasks suffer immensely. For example, when generating black-and-white sketches, the mid-tone bias causes outputs to appear grayish and blurry rather than crisp and binary (He et al.).
  • Video Generation: In video synthesis, the residual signal prevents the model from establishing a coherent global structure across frames from the very first step. This leads to low contrast, a washed-out appearance, and severe temporal inconsistencies or flickering (Shi et al.; Girdhar et al.).
  • Audio and Vocoders: In audio synthesis, the signal leak manifests as a persistent noise floor. Because the model's noise schedule never reaches absolute zero, it cannot generate true silence. This leaves a constant background hiss or artifact floor during intermittent silent patches in speech or isolated sound extraction (Nguyen et al.; Hai et al.).

Enforcing a Zero-SNR Slate

The solution to this problem is to explicitly adjust the noise schedule so that the SNR reaches exactly zero at the final timestep.

By enforcing a true zero terminal SNR, the model is forced during training to synthesize global structure, macro-level composition, and overall luminance from an entirely uninformative state (Menon et al.). This simple correction allows audio models to achieve absolute silence (Hai et al.), prevents over-exposure and artifacts in video (Gupta et al.), and empowers image models to generate full-contrast scenes with deep blacks and bright whites (Liu et al.).

While enforcing zero terminal SNR fixes the model's output distribution, it introduces a new mathematical challenge: singularities at the endpoints ($t=0$ and $t=1$). When the schedule is pushed to absolute limits, standard epsilon-prediction ($\epsilon$-prediction) models collapse because the target becomes mathematically undefined or unstable at the boundaries.

To circumvent this, researchers typically adopt one of two strategies:

  1. Velocity Prediction ($v$-prediction): As argued by Liu et al., switching the training target to $v$-prediction mathematically stabilizes the network at zero SNR, allowing it to smoothly predict a mixture of the data and the noise.
  2. Dedicated Boundary Modules: Alternatively, frameworks like SingDiffusion (Zhang et al.) act as plug-and-play modules that specifically correct the sampling math at these singular points, allowing pre-existing models to gain the benefits of extreme contrast generation without collapsing.

The Inversion Trade-off

While zero terminal SNR schedules dramatically improve generative quality, they introduce a distinct trade-off for downstream tasks that rely on inversion (mapping a real image or video back into noise for editing).

As Yesiltepe et al. observed, the mathematical collapse at the absolute final diffusion step makes the forward path non-invertible. Because the final step is completely uninformative, attempting to deterministically invert a video or image back into its original noise representation fails. This makes it difficult for zero-terminal SNR models to perfectly preserve subject identity during complex editing tasks like dynamic view synthesis, highlighting an ongoing area of research in balancing generative dynamic range with structural editability.

Key Foundational Work

Common Diffusion Noise Schedules and Sample Steps are Flawed (2024) — contributed 4 concept elements to this direction.
Exploiting the Signal-Leak Bias in Diffusion Models (2024) — contributed 4 concept elements to this direction.
FreGrad: Lightweight and Fast Frequency-aware Diffusion Vocoder (2024) — contributed 3 concept elements to this direction.
History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_noise_floor_suppression_49e3a414,
  title = {noise floor suppression},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/noise-floor-suppression/rdr_49e3a414df5c49268d54a3fa5d84be5d},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 1, frozen text):

@misc{lacuna_noise_floor_suppression_49e3a414_v1,
  title = {noise floor suppression},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/noise-floor-suppression/rdr_49e3a414df5c49268d54a3fa5d84be5d/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake