CFG-Zero*: Improved Classifier-Free Guidance for Flow Matching Models

Weichen Fan¹, Amber Yijia Zheng², Raymond A. Yeh², Ziwei Liu^1✉

S-Lab, Nanyang Technological University¹      Department of Computer Science, Purdue University ^2
✉Corresponding Author.

Paper | Project Page | Demo | Demo for Ghibli style

---

🔥 Huggingface demo for Ghibli style generation supported by EasyControl.

⚡️ Huggingface demo now supports text-to-image generation with SD3 and SD3.5.

💰 Bonus tip: You can even use pure zero-init (zeroing out the prediction of the first step) as a quick test—if it improves your flow-matching model a lot, it may indicate that the model has not converged yet.

🧪 Usage Tip: Use both optimized-scale and zero-init together. Adjust the zero-init steps based on total inference steps — 4% is generally a good starting point.

🔥 Update and News

• [2025.4.14] HiDream is suppported now!
• [2025.4.14] 🔥 Supported by sdnext now!
• [2025.4.6] 📙 Supported by EasyControl now!
• [2025.4.4] 🤗 Supported by Diffusers now!
• [2025.4.2] 🙌 Mentioned by Wan2.1!
• [2025.4.1] Qwen2.5-Omni is suppported now!
• [2025.3.30] Hunyuan is officially supported now!
• [2025.3.29] Flux is officially supported now!
• [2025.3.29] Both Wan2.1-14B I2V & T2V are now supported!
• [2025.3.28] Wan2.1-14B T2V is now supported! (Note: The default setting has been updated to zero out 4% of total steps for this scenario.)
• [2025.3.27] 📙 Supported by ComfyUI-KJNodes now!
• [2025.03.26] 📙 Supported by Wan2.1GP now!
• [2025.03.25] Paper|Demo|Code have been officially released.

Community Works

If you find that CFG-Zero* helps improve your model, we'd love to hear about it!

Thanks to the following models for supporting our method!

• SD.Next
• EasyControl
• ComfyUI-KJNodes
• Wan2.1GP
• ComfyUI Noted that ComfyUI's implementation is different from ours.

📑 Todo List

• Wan2.1
  • 14B Text-to-Video
  • 14B Image-to-Video
• Hunyuan
  • Text-to-Video
• SD3/SD3.5
  • Text-to-Image
• Flux
  • Text-to-Image (Guidance-distilled version)
  • Lora
• CogView4
  • Text-to-Image
• Qwen2.5-Omni
  • Audio generation
• EasyControl
  • Ghibli-Style Portrait Generation
• HiDream
  • text2image pipeline

😲 Gallery

▶ Click to expand comparison Images

▶ Click to expand comparison GIFs

Installation

1. Create a conda environment and install PyTorch

Note: You may want to adjust the CUDA version according to your driver version.

conda create -n CFG_Zero_Star python=3.10
conda activate CFG_Zero_Star

#Install pytorch according to your cuda version
conda install pytorch==2.5.1 torchvision==0.20.1 torchaudio==2.5.1 pytorch-cuda=12.4 -c pytorch -c nvidia

2. Install dependencies

pip install -r requirements.txt
apt install -y ffmpeg

Local demo

Host a demo on your local machine.

python demo.py

Inference

1. Wan2.1

a. Text-to-Video Generation

Simply run the following command to generate videos in the output folder. Noted that the current version is using Wan-AI/Wan2.1-T2V-14B-Diffusers with the default setting.

Noted that zero-steps for wan2.1 T2V is set to 1 (first 2 steps, 4% of the total steps).

python models/wan/T2V_infer.py

All results shown below were generated using this script on an H100 80G GPU.

CFG	CFG-Zero*
Prompt: "A cat walks on the grass, realistic" Seed: 1322140014
CFG	CFG-Zero*
Prompt: "A dynamic interaction between the ocean and a large rock. The rock, with its rough texture and jagged edges, is partially submerged in the water, suggesting it is a natural feature of the coastline. The water around the rock is in motion, with white foam and waves crashing against the rock, indicating the force of the ocean's movement. The background is a vast expanse of the ocean, with small ripples and waves, suggesting a moderate sea state. The overall style of the scene is a realistic depiction of a natural landscape, with a focus on the interplay between the rock and the water." Seed: 1306980124
CFG	CFG-Zero*
Prompt: "A stylish woman walks down a Tokyo street filled with warm glowing neon and animated city signage. She wears a black leather jacket, a long red dress, and black boots, and carries a black purse. She wears sunglasses and red lipstick. She walks confidently and casually. The street is damp and reflective, creating a mirror effect of the colorful lights. Many pedestrians walk about." Seed: 1270611998
CFG	CFG-Zero*
Prompt: "The camera follows behind a white vintage SUV with a black roof rack as it speeds up a steep dirt road surrounded by pine trees on a steep mountain slope, dust kicks up from it’s tires, the sunlight shines on the SUV as it speeds along the dirt road, casting a warm glow over the scene. The dirt road curves gently into the distance, with no other cars or vehicles in sight. The trees on either side of the road are redwoods, with patches of greenery scattered throughout. The car is seen from the rear following the curve with ease, making it seem as if it is on a rugged drive through the rugged terrain. The dirt road itself is surrounded by steep hills and mountains, with a clear blue sky above with wispy clouds." Seed: 158241056

b. Image-to-Video Generation

Follow Wan2.1 to clone the repo and finish the installation, then copy 'models/wan/image2video_cfg_zero_star.py' in this repo to the Wan2.1 repo (Wan2.1/wan). Modify 'Wan2.1/wan/init.py': replace 'from .image2video import WanI2V' with 'from .image2video_cfg_zero_star import WanI2V'.

Note: For I2V, zero_init_steps is set to 0 [2.5% zero out] by default to ensure stable generation. If you prefer more creative results, you can set it to 1 [5% zero out], though this may lead to instability in certain cases.

python generate.py --task i2v-14B --size 832*480 --ckpt_dir ./Wan2.1-I2V-14B-480P --image examples/i2v_input.JPG --prompt "Summer beach vacation style, a white cat wearing sunglasses sits on a surfboard. The fluffy-furred feline gazes directly at the camera with a relaxed expression. Blurred beach scenery forms the background featuring crystal-clear waters, distant green hills, and a blue sky dotted with white clouds. The cat assumes a naturally relaxed posture, as if savoring the sea breeze and warm sunlight. A close-up shot highlights the feline's intricate details and the refreshing atmosphere of the seaside." --base_seed 0

All results shown below were generated using this script on an H100 80G GPU.

Input Image	CFG	CFG-Zero*
Prompt: "Summer beach vacation style, a white cat wearing sunglasses sits on a surfboard. The fluffy-furred feline gazes directly at the camera with a relaxed expression. Blurred beach scenery forms the background featuring crystal-clear waters, distant green hills, and a blue sky dotted with white clouds. The cat assumes a naturally relaxed posture, as if savoring the sea breeze and warm sunlight. A close-up shot highlights the feline's intricate details and the refreshing atmosphere of the seaside." Seed: 0
Input Image	CFG	CFG-Zero*
Prompt: "Summer beach vacation style. A white cat wearing sunglasses lounges confidently on a surfboard, gently bobbing with the ocean waves under the bright sun. The cat exudes a cool, laid-back attitude. After a moment, it casually reaches into a small bag, pulls out a cigarette, and lights it. A thin stream of smoke drifts into the salty breeze as the cat takes a slow drag, maintaining its nonchalant pose beneath the clear blue sky." Seed: 3549111921

2. Flux

a. Text-to-Image Generation

We used black-forest-labs/FLUX.1-dev for the following experiment. Since this model is guidance-distilled, we applied only zero-init from our CFG-Zero* method. All images below were generated with the same seed on an H100 80G GPU.

python models/flux/Guidance_distilled.py

CFG	CFG-Zero*
Prompt: "a tiny astronaut hatching from an egg on the moon." Seed: 105297965

b. Lora

We used black-forest-labs/FLUX.1-dev with different loras for the following experiment. Since this model is guidance-distilled, we applied only zero-init from our CFG-Zero* method. All images below were generated with the same seed on an H100 80G GPU.

python models/flux/infer_lora.py

CFG	CFG-Zero*
Prompt: "Death Stranding Style. A solitary figure in a futuristic suit with a large, intricate backpack stands on a grassy cliff, gazing at a vast, mist-covered landscape composed of rugged mountains and low valleys beneath a rainy, overcast sky. Raindrops streak softly through the air, and puddles glisten on the uneven ground. Above the horizon, an ethereal, upside-down rainbow arcs downward through the gray clouds — its surreal, inverted shape adding an otherworldly touch to the haunting scene. A soft glow from distant structures illuminates the depth of the valley, enhancing the mysterious atmosphere. The contrast between the rain-soaked greenery and jagged rocky terrain adds texture and detail, amplifying the sense of solitude, exploration, and the anticipation of unknown adventures beyond the horizon." Seed: 875187112 Lora: https://civitai.com/models/46080/death-stranding

3. Hunyuan

We used hunyuanvideo-community/HunyuanVideo for the following experiment. Since this model is guidance-distilled, we applied only zero-init from our CFG-Zero* method. All images below were generated with the same seed on an H100 80G GPU.

python models/hunyuan/t2v.py

CFG	CFG-Zero*
Prompt: "In an ornate, historical hall, a massive tidal wave peaks and begins to crash. A man is surfing, cinematic film shot in 35mm. High quality, high defination." Seed: 376559893

4. SD3

We used stabilityai/stable-diffusion-3.5-large for the following experiment. All images below were generated with the same seed on an H100 80G GPU.

python models/sd/infer.py

CFG	CFG-Zero*
Prompt: "A capybara holding a sign that reads Hello World" Seed: 811677707

5. Qwen2.5-Omni

Install dependencies for Qwen2.5-Omni

pip install git+https://github.com/huggingface/transformers@f742a644ca32e65758c3adb36225aef1731bd2a8
pip install qwen-omni-utils[decord]
pip install flash-attn --no-build-isolation

Easy inference with CFG-Zero*

CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6 python models/Qwen2.5/infer.py

The following audios are generated by the script:

CFG	CFG-Zero⋆
🔊 Click to download	🔊 Click to download

6. EasyControl

Ghibli-Style Portrait Generation, the zero-init steps is set to 1 for default, feel free to try other values.

python models/easycontrol/infer.py

Source Image	CFG	CFG-Zero*

7. Cogview4

python models/Cogview4/infer.py

CFG	CFG-Zero*
Prompt: "A vibrant cherry red sports car sits proudly under the gleaming sun, its polished exterior smooth and flawless, casting a mirror-like reflection. The car features a low, aerodynamic body, angular headlights that gaze forward like predatory eyes, and a set of black, high-gloss racing rims that contrast starkly with the red. A subtle hint of chrome embellishes the grille and exhaust, while the tinted windows suggest a luxurious and private interior. The scene conveys a sense of speed and elegance, the car appearing as if it's about to burst into a sprint along a coastal road, with the ocean's azure waves crashing in the background." Seed: 42

8. HiDream

Git clone HiDream, and replace hidream_pipeline with ours 'models/HiDream/pipeline.py'

Then modify 'zero_steps' according the total inference steps.

cd HiDream-I1
python ./inference.py --model_type full

CFG	CFG-Zero*
Prompt: "A cat holding a sign that says \"Hi-Dreams.ai\"." Seed: 0

Easy Implementation

You can use this script to easily apply our method to any flow-matching-based model.

def optimized_scale(positive_flat, negative_flat):
  # Calculate dot production
  dot_product = torch.sum(positive_flat * negative_flat, dim=1, keepdim=True)

  # Squared norm of uncondition
  squared_norm = torch.sum(negative_flat ** 2, dim=1, keepdim=True) + 1e-8

  # st_star = v_condˆT * v_uncond / ||v_uncond||ˆ2
  st_star = dot_product / squared_norm
  return st_star

# Get the velocity prediction
noise_pred_uncond, noise_pred_text = model(...)
positive = noise_pred_text.view(Batchsize,-1)
negative = noise_pred_uncond.view(Batchsize,-1)

# Calculate the optimized scale
st_star = optimized_scale(positive,negative)

# Reshape for broadcasting
st_star = st_star.view(batch_size, *([1] * (len(noise_pred_text.shape) - 1)))

# Perform CFG-Zero* sampling
if sample_step == 0:
  # Perform zero init
  noise_pred = noise_pred_uncond * 0.
else:
  # Perform optimized scale
  noise_pred = noise_pred_uncond * st_star + \
  guidance_scale * (noise_pred_text - noise_pred_uncond * st_star)

BibTex

@misc{fan2025cfgzerostar,
      title={CFG-Zero*: Improved Classifier-Free Guidance for Flow Matching Models}, 
      author={Weichen Fan and Amber Yijia Zheng and Raymond A. Yeh and Ziwei Liu},
      year={2025},
      eprint={2503.18886},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2503.18886}, 
}

🔑 License

This code is licensed under Apache-2.0. The framework is fully open for academic research and also allows any commercial usage.

Disclaimer

We disclaim responsibility for user-generated content. The model was not trained to realistically represent people or events, so using it to generate such content is beyond the model's capabilities. It is prohibited for pornographic, violent and bloody content generation, and to generate content that is demeaning or harmful to people or their environment, culture, religion, etc. Users are solely liable for their actions. The project contributors are not legally affiliated with, nor accountable for users' behaviors. Use the generative model responsibly, adhering to ethical and legal standards.