Excited to present two papers at #ICLR2025 in Singapore 🇸🇬! If you are attending, please drop by our posters to talk about diffusion models 🔥. Details in the thread 🧵👇.

RT @puneeshdeora: NeurIPS releasing reviews one by one.

Join millions who have switched to Grok.

RT @Phc14097930: Excited to share minHiWave, my reimplementation of the HiWave paper! Clean code, minimal setup, and ready to run. Check it….

RT @fly51fly: [LG] Guidance in the Frequency Domain Enables High-Fidelity Sampling at Low CFG Scales.S Sadat, T Vontobel, F Salehi, R M. We….

Please check out our discussion here if you would like to see more information regarding the paper.

Pretty exciting to see two community implementations of our Frequency-Decoupled Guidance (FDG) in ComfyUI! 🚀. Huge thanks to the contributors:. 🔗 🔗 Note: I haven’t personally tested the code for correctness.

RT @kwangmoo_yi: Preprint of today: Sadat et al., "Guidance in the Frequency Domain Enables High-Fidelity Sampling at Low CFG Scales" -- ht….

RT @kwangmoo_yi: Preprint of today: Vontobel et al., "HiWave: Training-Free High-Resolution Image Generation via Wavelet-Based Diffusion Sa….

RT @rajabi2001: [1/7]⚡️Check out our recent work — "Token Perturbation Guidance for Diffusion Models". A simple yet effective method based….

Thanks, AK, for sharing our work!. Also, feel free to check out our related paper, where we explore how frequency analysis can generally enhance image generation in diffusion models:

This one-page style makes it easy for reviewers to just say they're not convinced, or they might stay confused due to lack of space for further clarification.

Based on my experience with the #ICCV review process as both an author and a reviewer, I believe the discussion system used at NeurIPS/ICLR/ICML is significantly more effective than the one-page rebuttal approach of ICCV/CVPR.

Finally, we show that FDG is compatible with any conditional diffusion model using CFG, and it consistently outperforms CFG across various models, datasets, and metrics. For full details and experiments, please check out the paper:.

FDG also enables analysis of Guidance Interval and Autoguidance in the frequency domain, offering insights into why these methods outperform standard CFG.

FDG is training-free, easy to implement, and introduces no noticeable overhead during sampling. We thus consider it a plug-and-play alternative to CFG for conditional diffusion models.

Building on this, we propose frequency-decoupled guidance (FDG)—a new method that applies separate guidance scales to low and high frequencies of the CFG update. FDG effectively boosts image quality at lower CFG scales while avoiding the trade-offs of high CFG values by design.

We also quantitatively show how low- and high-frequency components affect generation as guidance scale increases. Our results support the fact that low frequencies should be guided more conservatively to avoid the drawbacks of CFG at higher guidance scales.

We observe that excessive low-frequency guidance reduces diversity and causes oversaturation, while high frequencies benefit from higher guidance scales. This explains why uniform scaling in CFG degrades quality at low scales and boosts detail—but harms diversity—at high scales.

We analyze the CFG update rule in the frequency domain and show how different frequency components influence the final generation. Specifically, we find that low frequencies govern overall image structure and prompt alignment, while high frequencies control fine details.

Classifier-free guidance (CFG) is a key component of modern conditional diffusion models. However, the mechanisms through which CFG enhances quality, detail, and prompt alignment remain poorly understood. This work aims to bridge this gap.

🚨Happy to share our new work "Guidance in the Frequency Domain Enables High-Fidelity Sampling at Low CFG Scales" . TLDR: Applying classifier-free guidance in the frequency domain boosts quality at low CFG scales, while avoiding the issues of high CFG scales by design. 🧵👇