Guidance methods

[francois-rozet]

Description

This issue tracks the guidance methods that have been requested and/or implemented in azula.guidance. You are welcome to request new guidance methods. If you wish to contribute to Azula, this list is a good place to start.

List of requested guidance methods

• Denoising Diffusion Restoration Model (DDRM) by Kawar et al. (2022)
• RePaint by Lugmayr et al. (2022)
• Denoising Diffusion Null-space Model (DDNM) by Wang et al. (2023)
• Diffusion Models for Plug-and-Play Image Restoration (DiffPIR) by Zhu et al. (2023)
• Diffusion Posterior Sampling (DPS) by Chung et al. (2023)
• Pseudoinverse-Guided Diffusion Model (PGDM) by Song et al. (2023)
• Tweedie Moment Projected Diffusion (TMPD) by Boys et al. (2023)
• Twisted Diffusions Sampling (TDS) by Wu et al. (2023)
• Moment Matching Posterior Sampling (MMPS) by Rozet et al. (2024)
• Universal Guidance for Diffusion Model (UGDM) by Bansal et al. (2023)
• Manifold Preserving Guided Diffusion (MPGD) by He et al. (2024)
• Posterior Sampling with Latent Diffusion (PSLD) by Rout et al. (2023)
• Second-order Tweedie sampler from Surrogate Loss (STSL) by Rout et al. (2024)
• Rectified Flow Inversion (RFI) by Rout et al. (2024)
• Divide-and-Conquer Posterior Sampling (DCPS) by Janati et al. (2024)
• Midpoint Guidance Posterior Sampling (MGPS) by Janati et al. (2024)
• Mixture-Guided Diffusion Model (MGDM) by Janati et al. (2025)
• Decoupled Annealing Posterior Sampling (DAPS) by Zhang et al. (2024)
• Diffusion Tree Sampling (DTS) by Jain et al. (2025)
• Free Hunch (FH) by Rissanen et al. (2025)

[francois-rozet]

Hey @YazidJanati, would you like to contribute your methods to the library? Would love your feeback on the API as well 😉

[YazidJanati]

Hey @francois-rozet ! Thanks for tagging me, we will definitely start working on adding our methods (and the more recent one MGDM) to the library. In our private repo we have also implemented PGDM, DiffPIR, RePaint, PSLD, TDS, FPS, so maybe we could add those too :)

I haven't had the chance yet to test the API yet, but will do asap!

Tagging @Badr-MOUFAD

[francois-rozet]

Thanks for the fast reply! Tell me if you need any help, but I am sure you will feel right at home with the API 😉

In our private repo we have also implemented PGDM, DiffPIR, RePaint, PSLD, TDS, FPS, so maybe we could add those too

That would be fantastic! I would love to setup a standardized benchmark as well, both for pixel space and latent space guidance. For the latter, I will try to add stable diffusion and Flux plugins.

[YazidJanati]

Yes, adding SD and Flux would be great!

With @Badr-MOUFAD we can also help with the benchmark (which would be a significant feat). For example, it would be good add toy experiments for which the posterior and scores are given in closed form.