Rethinking Diffusion Model in High Dimension (Paper)
Natural Inference is a straightforward and general inference framework that does not rely on any probability concepts. It unifies most mainstream inference methods, including at least:
- ✅ DDPM Ancestral Sampling
- ✅ DDIM
- ✅ SDE Euler
- ✅ ODE Euler
- ✅ Heun
- ✅ DPMSolver
- ✅ DPMSolver++
- ✅ DEIS
- ✅ Flow Matching Euler
- For reproducing the results in Section 3.2 - Weighted Sum Degradation Phenome, please refer to src/AnalyzeWeightedSumDegradation.py.
- For reproducing the results in Section 4.3 - Represent DDPM Ancestral Sampling with Natural Inference Framework, please refer to src/AnalyzeDDPMDDIM.py.
- For reproducing the results in Section 4.4 - Represent DDIM with Natural Inference Framework, please refer to src/AnalyzeDDPMDDIM.py.
- For reproducing the results in Section 4.5 - Represent Flow Matching Euler Sampling with Natural Inference Framework, please refer to src/AnalyzeFlowMatching.py.
- For reproducing the results in Section 4.6 - Represent High Order Samplers with Natural Inference Framework, please refer to src/AnalyzeDEIS.py and src/AnalyzeDPMSolver.py.
- For reproducing the results in Section 5.3 - A Way to Control Image Sharpness, please refer to src/SD3NaturalInference.py.
- For reproducing the results in Section 5.4 - Better Coefficient Matrix, please refer to src/CIFAR10NaturalInference.py.
- To Validating the consistency between the results of the original algorithm and Natural Inference, please refer to ValidateNaturalInference.py.
- To visualizing coefficient matrixs on Natural Inference, please refer to Visualize/VisualizeCoeffMatrix.html. VisualizeCoeffMatrix.html is a standalone web application that can run offline in a browser. You can also open this link directly.