CFG scale(guidance scale) for generation

[CiaoHe]

Hi, when playing with the demo live, I found increase the default CFG(scale) will enhance the generated samples quality (less variance and shaper, like increase it to 10+)

Just curious about this

[ruoshiliu]

Yes, that’s an expected behavior of tuning classifier free guidance scale of diffusion models. Higher cfg scale produces less diversity and higher consistency with input observation, which is not necessarily a good thing depending on application. For single view novel view synthesis, which is a severely under constrained task, similar to text to image generation, diffusion model as a probabilistic image to image translation model is particularly apt as an architecture.