1. Pre-trained Models

• classifier_guidance-cifar10.pth

2. Samples

classifier_scale=30.0	classifier_scale=200.0

• The classes are "airplane", "automobile", "bird", "cat", "deer", "dog", "frog", "hose", "ship" and "truck" from top to bottom.

3. Theoretical Background

$$x_{t - 1} \leftarrow \text{sample from } \mathcal{N}(\mu + s\Sigma\nabla_{x_{t}}\log{p_{\phi}}(y \vert x), \Sigma)$$ $$\hat{\epsilon} \leftarrow \epsilon_{\theta}(x_{t}) - \sqrt{1 - \bar{\alpha}{t}}\nabla{x_{t}}\log{p_{\phi}}(y \vert x)$$ $$x_{t - 1} \leftarrow \sqrt{\bar{\alpha}{t - 1}}\Bigg(\frac{x{t} - \sqrt{1 - \bar{\alpha}{t}}\hat{\epsilon}}{\sqrt{\bar{\alpha}{t}}}\Bigg) + \sqrt{1 - \bar{\alpha}_{t - 1}}\hat{\epsilon}$$

4. To Do

• AdaGN
• BiGGAN Upsample/Downsample
• Improved DDPM sampling
• Conditional/Unconditional models
• Super-resolution model
• Interpolation

5. References

• [1] https://github.com/openai/guided-diffusion
• [2] https://github.com/openai/improved-diffusion