Fig. 1. Overview of generative adversarial network (GAN) principle

Fig. 2. Overview of variational auto-encoder (VAE)

Fig. 3. Overview of flow-based generative model

Fig. 4. Overview of diffusion model

Fig. 5. Comparison of different generative models

Fig. 6. Overview of image and video generation models

Fig. 7. Class-conditioned image generation^[95]

Fig. 8. Text-conditioned image generation^[96]

Fig. 9. Text-to-image generation results of Stable Diffusion^[97]

Fig. 10. Weight-tuning-based image customization^[102]

Fig. 11. Token-learning-based image customization^[103]

Fig. 12. Image customization with multi-concept composition^[104]

Fig. 13. Mask-region-based text-to-image editing^[105]

Fig. 14. Prompt-editing-based text-to-image editing^[106]

Fig. 15. Embedding-interpolation-based text-to-image editing^[107]

Fig. 16. Generated videos of Imagen Video^[108]

Fig. 17. VIDM framework using two diffusion model to generate video content and action information respectively^[111]

Fig. 18. PVDM framework that represents the video as three two-dimensional hidden variables, and thus uses the two-dimensional diffusion model for training^[112]

Fig. 19. Text-to-video generation results of Make-A-Video^[113]

Fig. 20. VideoFusion framework that uses pre-trained text-to-image diffusion model to generate base frame and uses video data to train a residual noise generator^[114]

Fig. 21. Video editing based on input image or text prompt^[115]

Fig. 22. One shot text-to-video generation^[116]