📢 Diffusion models (DM) generate samples from noise distribution, but for tasks such as image-to-image translation, one side is no longer noise. We present Denoising Diffusion Bridge Models, a simple and scalable extension to DMs suitable for distribution translation problems.

We extend diffusion processes to arrive at given endpoints, and propose “denoising bridge score matching“ to learn the intractable score for reversing this bridge process. We can exactly sample noisy views given data. The resulting algorithm is simple and scalable.

Our model can theoretical subsume many prior frameworks. We additionally propose improved network parameterization and hybrid SDE+ODE sampler for SOTA performance. w/ @aaron_lou @samar_a_khanna @StefanoErmon Paper url: Code:

seems like a pixel-aligned controlnet? but instead of a plug-in, this needs training for every catogories: canny, pose, depth, normal and etc?

Our framework focuses more on the development of the theory and is orthogonal to ControlNet, which is an architecture fine-tuning pretrained diffusion. Our framework can also use ControlNet architecture and fine-tune from pretrained diffusion weights.

@memdotai mem it

Excellent work！ Does your method require paired data to train？ Another questions，is h-transform critical in order to achieve the marginal distribution？If so，is the choice of the multiplier w in Eqn.13 sensitive to the data distribution for such requirement？