ComfyUI-LTX-FDG

Frequency-Decoupled Guidance (FDG) for ComfyUI and ComfyUI-LTXVideo.

Credit to dorpxam for bringing it up.

He might have more tested version soon https://github.com/AkaneTendo25/musubi-tuner/issues/40#issuecomment-4033048905

Vibecoded based on the paper:

"Guidance in the Frequency Domain Enables High-Fidelity Sampling at Low CFG Scales" https://arxiv.org/abs/2506.19713

What is FDG?

FDG improves upon standard Classifier-Free Guidance (CFG) by applying separate guidance scales to different frequency components:

• Low-frequency guidance (w_low): Controls global structure, composition, and color

  • Lower values preserve diversity and prevent oversaturation
  • Higher values improve condition alignment

• High-frequency guidance (w_high): Enhances visual fidelity, details, and sharpness

  • Can use higher values without the drawbacks of high CFG

Installation

1. Clone or download this repository into your ComfyUI custom_nodes folder:

   ComfyUI/custom_nodes/ComfyUI-LTX-FDG/
   

2. Restart ComfyUI

3. The FDGParameters node will appear in the lightricks/LTXV category

Usage

Basic Workflow with ComfyUI-LTXVideo

┌─────────────────────┐
│  FDGParameters      │
├─────────────────────┤
│ modality: VIDEO     │
│ cfg: 1.5            │
│ fdg_enabled: ON     │
│ w_low: 1.5          │
│ w_high: 5.0         │
│ frequency_levels: 2 │
└─────────┬───────────┘
          │
          └─> parameters ──> MultimodalGuider ──> SamplerCustomAdvanced

Video-Only (LTXV models):

┌─────────────┐     ┌─────────────┐     ┌──────────────┐
│  Positive   │────>│ Multimodal  │────>│   Sampler    │
│  Negative   │────>│  Guider     │     │   Custom     │
│  Model      │────>│  (LTXV)     │     │   Advanced   │
└─────────────┘     │             │     └──────────────┘
┌─────────────┐     │             │
│ FDGParams   │────>│             │
│ (VIDEO)     │     │             │
└─────────────┘     └─────────────┘

Video+Audio (LTXAV models):

Use two FDGParameters nodes (one for VIDEO, one for AUDIO):

┌──────────────────┐     ┌──────────────────┐
│ FDGParameters    │     │ FDGParameters    │
│ modality: VIDEO  │     │ modality: AUDIO  │
│ w_low: 1.5       │     │ w_low: 1.0       │
│ w_high: 5.0      │     │ w_high: 3.0      │
└────────┬─────────┘     └────────┬─────────┘
         │                       │
         └───────┬───────────────┘
                 ▼
         ┌───────────────┐
         │  Multimodal   │
         │    Guider     │
         │  (LTXAV)      │
         └───────────────┘

Recommended Settings

Video-Only (LTXV)

| Parameter | Value | Description | |-----------|-------|-------------| | cfg | 1.5 | Low CFG for better diversity | | w_low | 1.0-1.5 | Conservative for temporal stability | | w_high | 4.0-7.0 | Higher for sharp details | | frequency_levels | 2 | Default, good balance |

Video+Audio (LTXAV)

Video:

• w_low: 1.0-1.5 - Prevent temporal flicker
• w_high: 4.0-7.0 - Sharp video details

Audio:

• w_low: 1.0 - Maintain audio quality
• w_high: 2.0-3.0 - Moderate enhancement

Paper's Recommended Settings (Table 8)

| Model Type | cfg | w_low | w_high | Use Case | |------------|-----|-------|--------|----------| | SDXL | 3.0 | 3.0 | 10.0 | High quality | | SD3 | 1.5 | 1.5 | 12.0 | Maximum detail | | EDM2-XL | 2.0 | 1.0 | 2.0 | Fast generation |

Key Benefits

1. Better quality at low CFG scales - Get high-detail results without high CFG
2. Reduced oversaturation - Avoid color artifacts from high guidance scales
3. Maintained diversity - Keep the natural variation of low CFG sampling
4. Plug-and-play - No model retraining needed

Performance

FDG adds computational overhead:

| Setting | Relative Speed | |---------|----------------| | Standard CFG | 1x (baseline) | | FDG 2 levels | ~1.3-1.5x slower | | FDG 3+ levels | ~1.5-2x slower |

To speed up, reduce frequency_levels to 1.

How It Works

FDG uses Laplacian pyramids to decompose the model prediction into frequency bands:

1. Build Laplacian pyramid from conditional prediction
2. Build Laplacian pyramid from unconditional prediction
3. Apply different guidance scales to each frequency level
4. Reconstruct from guided pyramid

Troubleshooting

Q: The generated videos look oversaturated A: Reduce w_low to 1.0-1.5

Q: Details are blurry A: Increase w_high to 5.0-10.0

Q: Too much flicker between frames A: Reduce both w_low and w_high for more temporal consistency

Q: FDG doesn't seem to work A: Make sure fdg_enabled is ON in FDGParameters and the parameters are connected to MultimodalGuider

Q: Sampling is slower with FDG A: This is expected. To speed up, reduce frequency_levels to 1, or disable FDG with fdg_enabled: OFF

Q: Error with cfg=1.0 A: FDG requires cfg > 1.0 to work. Use cfg=1.5 or higher.

Citation

If you use FDG in your research, please cite:

@article{sadat2025fdg,
  title={Guidance in the Frequency Domain Enables High-Fidelity Sampling at Low CFG Scales},
  author={Sadat, Seyedmorteza and Vontobel, Tobias and Salehi, Farnood and Weber, Romann M.},
  journal={arXiv preprint arXiv:2506.19713},
  year={2025}
}

License

This implementation follows the same license as ComfyUI.

References

• Paper: https://arxiv.org/abs/2506.19713
• Laplacian Pyramids: Burt & Adelson (1983)
• ComfyUI-LTXVideo: https://github.com/Lighttricks/ComfyUI-LTXVideo