GSoC 2026: Question about Project 1 & 3

[Jocho-Smith]

Hi sbi maintainers,

Janos Gabler suggested I reach out regarding GSoC and potential projects in sbi, so I wanted to introduce myself and ask for your guidance.

I reviewed the 2026 project ideas and noticed a strong focus on strengthening core inference infrastructure while modernizing methods and APIs. I am particularly interested in:

• Project 1 – Integrate Azula Library for Diffusion Sampling
• Project 3 – Potential Function API Redesign

At this stage, I am undecided which direction would be the better strategic fit, and I would value your perspective.

For context: I am a PhD candidate in applied mathematics with some training in Bayesian statistics. I am comfortable with Python, PyTorch, SciPy/Numpy and git/GitHub, and I regularly work in research codebases.

Two honest gaps:

• For Project 1, I do not yet have hands-on experience with score-based generative models, only conceptually about VAEs, normalizing flows and GANs.
• For Project 3, my exposure to advanced MCMC and gradient-based sampling concepts is currently more general/theoretical than implementation-heavy.

That said, I am prepared to learn quickly in either direction.

I have a slight preference for Project 1 because the integration work appears more 'technically implementation-heavy', which aligns well with my interests for GSoC. However, I also find the API redesign in Project 3 appealing.

Which of the two would you recommend I focus on? And from your perspective, what distinguishes a strong GSoC proposal in either direction?

In the meantime I'll look for a beginner-friendly issue :D

Best regards and thanks for the nice project ideas!

[janfb]

Hi @Jocho-Smith

thanks for reaching out, this sounds great!

Given your background I would tend towards Project 3 for you. It's an interesting project because the potential_fn is used in many different scenarios and you will likely learn a lot about different sbi methods. Still, I see the project more on the software engineering / design side, less on the algorithmic side. Of course, one needs to understand how the potential are being used, but this project is not about implementing a new method or feature.

For preparation, I suggest you have a look into the core codebase to understand how we use the potential functions (mostly to enable different sampling methods like MCMC, VI, ...). Along the way you will likely notice that the API is a bit confusing and clunky and improving this is the idea of this project :)
There is also a recent issue #1783 related to this and to the way how we initialize the potential.

Regarding a strong proposal, I would consult the GSoC Guidelines for structuring it and the rest comes from understanding the core problem and having a couple of meaningful suggestions how to fix. But we are happy to help with that along the way!

Let me know if you have any questions!

Cheers,
Jan

[Jocho-Smith]

Hi @janfb,

thanks for the clarification. I'll gladly focus on Project 3. So my next steps will be:

• understanding how sbi uses the potential_fn method
• do some minor contribution along the way by contributing to the the discussion of issue #1783 , since your guidelines asked for "making a contribution to the sbi codebase" here. Or do you recommend something else for the beginning?
• consolidating the GSoC guidelines for the proposal

Where do you prefer the following communication to happen? Also, do you have someone already in mind for mentoring this project and is it ok, if I sparsely contact this person (or you) the next days directly for questions?

Best,
Jocho

[manuelgloeckler]

Hey @Jocho-Smith ,

Yes doing a minor contribution would be a next step (please also one that is solving one of the current issues). There are also some issues labeled as "good first issue", for any new problem please first create an issue.

I would be available for answering questions the next days. At best join the discord server (i.e. here [edit: nvm you just joined already]), you can PN me (user name manuel_gloeckler) if you have any question.

Kind regards,
Manuel