Hadar Davidson

[Hadar Davidson @HadarDavidson]

Excited to share Colored Noise Sampling (CNS)!🎉
Instead of injecting white noise, our SDE sampler exploits the inherent spectral bias of diffusion models. We dynamically color the injected noise to focus on frequencies where details are missing, substantially improving FID.🧵1/9

[Noam Issachar @IssacharNoam]

1/6 Diffusion models are scaling up, but deploying a massive, monolithic network uniformly across the entire generative timeline is inherently inefficient.
Introducing Complexity-Balanced Splitting (CBS): a principled framework that allocates capacity exactly where needed!👇🧵

[toyxyz @toyxyz3]

https://t.co/cRmdEPtJbV

[Hadar Davidson @HadarDavidson]

@BrianCChao @LucaAmb In contrast, CNS is a new *sampler*. We modify the standard SDE by injecting colored noise (based on expected current spectral SNR) at each step instead of white noise. No retraining! We also show substantial improvements using CNS even alongside different forward processes. 🚀

[Hadar Davidson @HadarDavidson]

Excited to share Colored Noise Sampling (CNS)!🎉
Instead of injecting white noise, our SDE sampler exploits the inherent spectral bias of diffusion models. We dynamically color the injected noise to focus on frequencies where details are missing, substantially improving FID.🧵1/9

[Hadar Davidson @HadarDavidson]

@BrianCChao @LucaAmb Thanks! Great question (we actually dive into this in the Related Work + Experiments😁).

EqualSNR proposes an alternative forward process to keep SNR equal across all frequencies, mitigating spectral bias. Crucially, this approach requires training a model specifically for it.

[Hadar Davidson @HadarDavidson]

9/9 - 🔗 Links
🌐 Project page: https://t.co/bLkUPqpADA
📄 Paper: https://t.co/y35yFM67iB
💻 Code: https://t.co/u3QQlg5afC
🤗 HF: https://t.co/L16YDSXPB3
Joint work with @IssacharNoam & @BenaimSagie

[Hadar Davidson @HadarDavidson]

8/9 - Step-efficient inference⚡
CNS at 100 steps already beats standard SDE at 1000 steps (FID 7.2 vs 7.8).
It is robust across solver orders (Euler, Heun, SRK2), and slots into FLUX.1-dev and FLUX.2-klein for text-to-image. All at inference time.