BDIA-DDIM Scheduler

Model/Pipeline/Scheduler description

The BDIA-DDIM scheduler was first applied in stable diffusion in the ECCV 2024 paper "Exact Diffusion Inversion via Bi-directional Integration Approximation" by Guoqiang Zhang, J. P. Lewis, and W. Bastiaan Kleijn. Below are results from the initial implementation for text-to-image generation using StableDiffusion V2.

The BDIA-DDIM Scheduler improves image sampling quality over traditional DDIM through a more accurate integral approximation, with negligible computational overhead. Experiments further show that BDIA-DDIM produces markedly better image sampling qualities than DDIM for text-to-image generation, thanks to the more accurate integration approximation. Since the first implementation, I have forked the diffusers library to integrate BDIA-DDIM as an additional scheduler and have found it produces significantly better results, especially at lower timesteps.

Open source status

• The model implementation is available.
• The model weights are available (Only relevant if addition is not a scheduler).

Provide useful links for the implementation

Original Paper: "Exact Diffusion Inversion via Bi-directional Integration Approximation" by Guoqiang Zhang, J. P. Lewis, and W. Bastiaan Kleijn.

Original Implementation: https://github.com/guoqiang-zhang-x/BDIA

My diffusers implementation: https://github.com/Jdh235/diffusers

Beginne damit, das BDIA-DDIM-Paper und die ursprüngliche BDIA-Implementierung zu lesen, und vergleiche dann den verlinkten diffusers-Fork mit dem aktuellen Repository. Ermittle, wie der Scheduler neben den vorhandenen Stable Diffusion-Schedulern eingeordnet werden sollte, und stelle sicher, dass sein Sampling-Verhalten abgedeckt ist, bevor du das Feature als vollständig betrachtest.