REVIEW 6 cited by
Alleviating Exposure Bias in Diffusion Models through Sampling with Shifted Time Steps
Not yet reviewed by Pith; the record is open.
This paper has not been read by Pith yet. Machine review is queued; the pith claim, tier, and objections will appear here once it completes.
SPECIMEN: schema-true, not a live event
T0 review · schema-true
One-sentence machine reading of the paper's core claim.
pith:XXXXXXXX · record.json · timestamp
abstract
Diffusion Probabilistic Models (DPM) have shown remarkable efficacy in the synthesis of high-quality images. However, their inference process characteristically requires numerous, potentially hundreds, of iterative steps, which could exaggerate the problem of exposure bias due to the training and inference discrepancy. Previous work has attempted to mitigate this issue by perturbing inputs during training, which consequently mandates the retraining of the DPM. In this work, we conduct a systematic study of exposure bias in DPM and, intriguingly, we find that the exposure bias could be alleviated with a novel sampling method that we propose, without retraining the model. We empirically and theoretically show that, during inference, for each backward time step $t$ and corresponding state $\hat{x}_t$, there might exist another time step $t_s$ which exhibits superior coupling with $\hat{x}_t$. Based on this finding, we introduce a sampling method named Time-Shift Sampler. Our framework can be seamlessly integrated to existing sampling algorithms, such as DDPM, DDIM and other high-order solvers, inducing merely minimal additional computations. Experimental results show our method brings significant and consistent improvements in FID scores on different datasets and sampling methods. For example, integrating Time-Shift Sampler to F-PNDM yields a FID=3.88, achieving 44.49\% improvements as compared to F-PNDM, on CIFAR-10 with 10 sampling steps, which is more performant than the vanilla DDIM with 100 sampling steps. Our code is available at https://github.com/Mingxiao-Li/TS-DPM.
Forward citations
Cited by 6 Pith papers
-
DMQ: Dissecting Outliers of Diffusion Models for Post-Training Quantization
A post-training quantization method that combines learned channel scaling and power-of-two scaling keeps diffusion image quality high at 4-bit weight, 6-bit activation precision.
-
FreeDNA: Endowing Domain Adaptation of Diffusion-Based Dense Prediction with Training-Free Domain Noise Alignment
Aligning the L2-norm statistics of noise predictions during diffusion sampling improves domain adaptation for dense prediction, with a source-free version guided by high-confidence regions.
-
Explorative Modeling: Unlocking a Third Pretraining Axis and End-to-End Generation
Training on the best of K generated candidates improves image, video, and language generative models, with the reported gains growing with scale and enabling single-pass end-to-end generation.
-
Integration Matters: Rollout-Based Training for Constrained Diffusion Models
Rollout-based fine-tuning with a learned adaptive guidance scaling yields near-zero constraint violations while preserving sample fidelity in constrained diffusion models.
-
DiffIER: Optimizing Diffusion Models with Iterative Error Reduction
DiffIER claims that iteratively minimizing the distance between a diffusion model's predicted noise and a random Gaussian sample at each inference step improves generation quality.
-
Pretrained Diffusion Models Are Inherently Skipped-Step Samplers
A DDPM-trained noise predictor can denoise across several time steps in one update because the multi-step posterior is Gaussian and uses the same network.
Discussion (0). Continue with ORCID to comment.