On the Importance of Noise Scheduling for Diffusion Models

Chen, Ting

doi:10.48550/arxiv.2301.10972

Cited by 7 publications

(5 citation statements)

References 9 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…(3) t and T represent the t th diffusion step and the total number of diffusion steps, respectively. Following the experiments described in [44] we analyzed the results obtained when τ is 1, 2, and 3. The results of this experiment are shown in Table II.…”

Section: Resultsmentioning

confidence: 99%

Synthetic Echocardiograms Generation Using Diffusion Models

Pellicer,

Singh Yadav,

Bhagtani

et al. 2023

Preprint

View full text Add to dashboard Cite

An echocardiogram is a video sequence of a human heart captured using ultrasound imaging. It shows heart structure and motion and helps in diagnosis of cardiovascular diseases. Deep learning methods, which require large amounts of training data have shown success in using echocardiograms to detect cardiovascular disorders such as valvular heart disease. Large datasets of echocardiograms that can be used for machine learning training are scarce. One way to address this problem is to use modern machine learning generative methods to generate synthetic echocardiograms that can be used for machine learning training. In this paper, we propose a video diffusion method for the generation of echocardiograms. Our method uses a 3D selfattention mechanism and a super-resolution model. We demonstrate that our proposed method generates echocardiograms with higher resolution and with lesser artifacts, compared to existing echocardiogram generation methods.

show abstract

Section: Resultsmentioning

confidence: 99%

Synthetic Echocardiograms Generation Using Diffusion Models

Pellicer,

Singh Yadav,

Bhagtani

et al. 2023

Preprint

View full text Add to dashboard Cite

show abstract

“…Several hyperparameters control the diffusion process in kdiffusion, and we extensively explore the role of training noisedistributed according to a log-normal distribution with parameters (Pmean, P std ) -and sampling noise with boundary values of σmin and σmax. These distributions are crucial choices depending on the task and on the dataset [32]. Given we use diffusion models on an The reverse process of a DM does not need to start from noise with variance σ 2 max but it can place at any arbitrary step t ∈ (0, T ), with σ 2 max = σ 2 0 as shown in [26].…”

Section: Methodsmentioning

confidence: 99%

Exploring Diffusion Models for Unsupervised Video Anomaly Detection

Tur,

Dall’Asen,

Beyan

et al. 2023

2023 IEEE International Conference on Image Processing (ICIP)

View full text Add to dashboard Cite

This paper investigates the performance of diffusion models for video anomaly detection (VAD) within the most challenging but also the most operational scenario in which the data annotations are not used. As being sparse, diverse, contextual, and often ambiguous, detecting abnormal events precisely is a very ambitious task. To this end, we rely only on the information-rich spatio-temporal data, and the reconstruction power of the diffusion models such that a high reconstruction error is utilized to decide the abnormality. Experiments performed on two large-scale video anomaly detection datasets demonstrate the consistent improvement of the proposed method over the state-of-the-art generative models while in some cases our method achieves better scores than the more complex models. This is the first study using a diffusion model and examining its parameters' influence to present guidance for VAD in surveillance scenarios.

show abstract

“…Our action sequence prediction approach partially mitigates this issue, but may not suffice for tasks requiring high rate control. Future work can exploit the latest advancements in diffusion model acceleration methods to reduce the number of inference steps required, such as new noise schedules [7], inference solvers [23], and consistency models [47].…”

Section: Limitations and Future Workmentioning

confidence: 99%