FreSca is a novel frequency scaling approach that enhances diffusion models without additional training. Our key findings:

• Frequency scaling improves visual quality with minimal computational overhead
• Energy-based cutoff strategy provides stable and consistent improvements
• Compatible with existing diffusion models as a plug-and-play enhancement (image generation, video generation, depth estimation, image editing)

• [2025-05] 🔥 Major Update! Our new version reveals the optimal locations for frequency scaling, showing improvements across image generation, video synthesis, depth estimation, and editing tasks. New energy-based cutoff technique delivers superior stability!
• [2025-04] Exciting discovery! FreSca boosts video diffusion models' performance without training! Preliminary VideoCrafter2 results added.
• [2025-04] Released example implementation for FreSca.

VideoCrafter2: Original vs FreSca
Original
FreSca
"A motorcycle riding along a desert highway, sand dunes stretching beside"	"Lanterns drifting into the night sky, a calm lake mirroring their glow"	"A train winding through sunflower fields, bright yellow blooms on both sides"	"A campfire on a lakeshore, stars sparkling in the dark sky"	"A tiger walks in the forest, photorealistic, 4k, high definition"

Click on any GIF to view the full MP4 video.

FreSca enhances depth estimation quality by preserving fine details and improving edge accuracy.

The core implementation is available in core/fresca.py and can be easily integrated into any diffusion model.

1. Frequency Space Transformation: Convert latent features to frequency domain using FFT
2. Selective Scaling: Apply customizable scaling to low- or high-frequency components
3. Energy-based Cutoff: Adaptive boundary determination for frequency scaling

• Cutoff ratio: 0.5~0.9 (energy-based approach)
• High-frequency scaling factor: 1.0 < h < 1.5
• Position: Apply the nosie predictions

Find more example implementations in the demo/ directory.

If you use this code for your research, please cite our work:

@article{huang2025fresca,
        title={FreSca: Unveiling the Scaling Space in Diffusion Models},
        author={Huang, Chao and Liang, Susan and Tang, Yunlong and Ma, Li and Tian, Yapeng and Xu, Chenliang},
        journal={arXiv preprint arXiv:2504.02154},
        year={2025}
}

For questions or feedback, please open an issue on GitHub or contact the authors directly.