Is Diffusion Really Stronger Than GAN?

AI painting is one of the branches of AIGC, amidst the hype and controversy, in 2022 it was even dubbed the “AIGC Year.” As AI painting exploded, one of its core technologies, Diffusion Model, has also gained immense popularity in the field of image generation, even showing signs of starting to surpass GAN.

From the principle diagram, it can be seen that the input text is first encoded, then converted into a 64*64 small image by a text-to-image diffusion model, and then processed by a super-resolution diffusion model to enhance the image resolution during further iterations, resulting in the final generated image—a 1024*1024 image.

Scan the code to receive essential papers on diffusion

The biggest advantage of the diffusion model is its simple training. It utilizes UNet from the field of image segmentation, resulting in stable training loss and excellent model performance. Compared to GAN, which requires adversarial training with a discriminator, or VAE that needs variational posterior, the loss of the diffusion model is truly much simpler. The diffusion model only needs to “mimic” the inverse process corresponding to a very simple forward process. This straightforward and efficient training also allows the diffusion model to perform exceptionally well in many tasks, even surpassing GAN.

Overall, the diffusion model field is in a flourishing state, somewhat reminiscent of the time when GAN was first introduced. However, the current training techniques allow the diffusion model to directly skip the model tuning phase in the GAN field and can be used directly for downstream tasks. There are still some core theoretical issues in this field that need research, providing valuable research content for scientific practitioners, sparking many ideas. Moreover, since this model already works well, its integration with downstream tasks is only just beginning, presenting numerous opportunities to quickly secure a place. In the future, as the existing problems in diffusion models are solved, diffusion models will gradually dominate the field of deep generative models.

Recently, during a conversation with a prominent conference paper author, I was amazed by his background. This expert has already published over twenty top conference papers! (Definitely at the level of a top conference paper collector). Besides being a reviewer for conferences and journals like AAI, CVPR, MICCAI, ECCV, ACCV, and Pattern Recognition, he also worked as an Applied Scientist at Amazon. Each of these descriptions is not simple~

On February 7, this top conference paper author will be invited to talk about the diffusion model, to see how the expert understands diffusion.

Scan the code to reserve a free live course with the top conference expert

Limited to the first hundred fans

Benefits at the end of the article

To conduct research efficiently, mastering methods and reasonably utilizing available resources is crucial! Based on my experience with junior peers and my own research growth, I found that everyone generally lacks a systematic research knowledge system, making it difficult to write a qualified article. Not to mention discovering good innovative points and ideas!

Research and paper writing can be categorized into three types based on outcomes:

1. Discovering a new method and applying it to a known problem(Old Problem New Method)

2. Discovering a new problem and expanding an existing study to this problem(New Problem Old Method)

3. Discovering a new problem and proposing a new method to analyze it(New Problem New Method)

In terms of difficulty, New Problem Old Method < New Problem New Method < Old Problem New Method.

Therefore, the innovative points and ideas that everyone finds troublesome actually have methodologies. Once mastered, it becomes relatively easier to find an idea～

To generate ideas, there are two treasures—models + algorithms

Models can be understood as parts of the system model and can be extensions of existing problem models. In fact, this is called innovation; as long as what you do has not been studied or is studied very little by others, it can already be considered innovation.

When you propose a model, you need to choose a suitable algorithm to solve the problems within the model. It can involve finding a suitable algorithm and reasonably improving it according to your scenario, which is already a great idea.

Scan the code to reserve a free live course with the top conference expert

Limited to the first hundred fans

Benefits at the end of the article