Open-source temporal denoiser for Blender animations - looking for feedback

Hi everyone,

We’re currently developing an open-source temporal denoiser for Blender Cycles focused on animation rendering. The goal is to reduce the huge sample counts currently needed to get stable, flicker-free animations.

The model is trained on large amounts of Cycles renders using auxiliary passes (albedo, normals, depth, etc.) and high-quality reference frames. We’re building the full pipeline ourselves: dataset generation, preprocessing, validation, and training.

The dataset is the first denoising dataset to be opensource and high quality.

Current denoisers work reasonably well for still images, but temporal stability in animations is still a major issue, especially at low sample counts. We’re trying to explore whether a dedicated temporal model could significantly improve this while remaining practical for Blender integration.

We thus came up with a reccurent Unet architecture for fast and stable denoising that work especialy good with cycle.

We’d really like feedback from the community:

What are the biggest issues you currently face with animation denoising? What would you expect from a good temporal denoiser? Would integration directly into Cycles be something people would use?

Any thoughts, concerns, or ideas are welcome.