Non-Cross Diffusion for Semantic Consistency
Abstract: In diffusion models, deviations from a straight generative flow are a common issue, resulting in semantic inconsistencies and suboptimal generations. To address this challenge, we introduce Non-Cross Diffusion, an innovative approach in generative modeling for learning ordinary differential equation (ODE) models. Our methodology strategically incorporates an ascending dimension of input to effectively connect points sampled from two distributions with uncrossed paths. This design ensures enhanced semantic consistency throughout the inference process, which is especially critical for applications reliant on consistent generative flows, including distillation methods and deterministic sampling, which are fundamental in image editing and interpolation tasks.
Our empirical results demonstrate the effectiveness of Non-Cross Diffusion, showing a substantial reduction in semantic inconsistencies at different inference steps and a notable enhancement in the overall performance of diffusion models.
Submission Length: Regular submission (no more than 12 pages of main content)
Changes Since Last Submission: All revised parts are highlighted in blue. Here are the details:
* We have revised Sec. 3.2 to further discuss why crossing will lead to *incorrect target*.
* We have added further discussion about trivial solutions with initial noise as condition in Sec. C.
* We revise the Training Stage in Sec. 3.3 to clarify why bootstrapping contributes to robustness.
* We have added more details in Implementation Details in Sec. 4.2 to clarfy that all models in our paper are training without label as condition.
* We have add Sec. D to the Appendix to futher discuss the xflow in real life model in term of existence of crossing and redirection with different inference step.
Assigned Action Editor: ~Charles_Xu1
Submission Number: 2512
Loading