Is Rectified Flow theoretically better than Diffusion Model?
Analyzing why diffusion models may underperform compared to rectified flow in practice, and how to "cook" a high-quality model.
Latest Posts
External Discussions
Why is Pre-norm more frequently used than Post-norm?
Conclusion: Pre-norm does not disturb the gradient and scales cleanly.
The shape regularity of diffusion sampling trajectories
Conclusion: Trajectories follow a linear-nonlinear-linear structure in a 3D sub-space.
3. Flexible patch size for ViT: How should we resize patch embedding?
Conclusion: Start from principled experiment, the authors propose PI-Resize (a combination of transformation matrixs derived from bilinear interpolation) for "losslessly" upsampling weight embedding.