AdaEdit: Adaptive Temporal and Channel Modulation for Flow-Based Image Editing

Guandong Li, Zhaobin Chu
iFLYTEK Research

Abstract

AdaEdit is a training-free adaptive image editing framework designed for Flow-based diffusion models. We address the "injection dilemma" in inversion-based editing through two core innovations:

(1) Progressive Injection Schedule: Replaces binary truncation with continuous decay functions (sigmoid/cosine/linear), eliminating feature discontinuities and reducing hyperparameter sensitivity.

(2) Channel-Selective Latent Perturbation: Applies strong perturbations to edit-relevant channels while maintaining weak perturbations on structural channels based on channel importance estimation, achieving effective editing while preserving structural fidelity.

On PIE-Bench (700 images), AdaEdit achieves significant improvements over baseline methods:

• LPIPS ↓ 8.7% (background preservation)
• SSIM ↑ 2.6% (structural similarity)
• PSNR ↑ 2.3% (peak signal-to-noise ratio)
• CLIP ≈ -0.9% (editing accuracy nearly lossless)

Installation

# Clone repository
git clone https://github.com/leeguandong/AdaEdit.git
cd AdaEdit

# Install dependencies
pip install -r requirements.txt

Model Download

Download FLUX.1-dev model weights and place them in the checkpoints/ directory:

# Download FLUX.1-dev model
huggingface-cli download black-forest-labs/FLUX.1-dev --local-dir checkpoints/flux-dev

Usage

Basic Usage

python adaedit.py \
    --source_img source.jpg \
    --source_prompt "A photo of a cat" \
    --target_prompt "A photo of a dog" \
    --output_dir outputs/

Full Parameters

python adaedit.py \
    -i source.jpg \
    -sp "A photo of a cat" \
    -tp "A photo of a dog" \
    -o outputs/ \
    --edit_object "cat" \
    --num_steps 15 \
    --guidance 4.0 \
    --inject 4 \
    --inject_schedule sigmoid \
    --kv_mix_ratio 0.9 \
    --ls_ratio 0.25 \
    --use_channel_ls \
    --channel_ls_temp 1.0 \
    --seed 42

Parameters

Parameter	Short	Default	Description
--source_img	-i	Required	Source image path
--source_prompt	-sp	Required	Source image description
--target_prompt	-tp	Required	Target editing description
--output_dir	-o	outputs/	Output directory
--edit_object		""	Object to edit (for mask extraction)
--num_steps		15	Number of sampling steps
--guidance		4.0	Guidance scale
--inject		4	Injection step threshold
--inject_schedule		sigmoid	Injection schedule: binary/sigmoid/cosine/linear
--kv_mix_ratio		0.9	KV-Mix ratio
--ls_ratio		0.25	Latents-Shift strength
--use_channel_ls		False	Enable channel-selective Latents-Shift
--channel_ls_temp		1.0	Channel importance temperature
--seed		0	Random seed (0=random)
--offload		False	Low VRAM mode

Examples

# Example 1: Object replacement (with progressive sigmoid schedule)
python adaedit.py \
    -i examples/cat.jpg \
    -sp "A photo of a cat on a sofa" \
    -tp "A photo of a dog on a sofa" \
    --edit_object "cat" \
    --inject_schedule sigmoid \
    --use_channel_ls

# Example 2: Style transfer (without Latents-Shift)
python adaedit.py \
    -i examples/portrait.jpg \
    -sp "A portrait photo" \
    -tp "A portrait in anime style" \
    --edit_type style \
    --inject_schedule cosine

# Example 3: Low VRAM mode
python adaedit.py \
    -i examples/scene.jpg \
    -sp "A street scene" \
    -tp "A street scene at night" \
    --offload

Key Features

1. Progressive Injection Schedule

Traditional methods use binary truncation (first N steps inject=1, subsequent steps inject=0), causing feature discontinuities. AdaEdit provides three continuous decay functions:

• Sigmoid (recommended): Smooth transition, moderate sharpness
• Cosine: Cosine decay, smoother
• Linear: Linear decay, simplest

2. Channel-Selective Latent Perturbation

Different channels encode different information (structure/color/texture). AdaEdit automatically estimates the edit-relevance of each channel:

• Edit-relevant channels: Strong perturbation (promotes content change)
• Structural channels: Weak perturbation (maintains layout stability)

3. Plug-and-Play

AdaEdit is training-free and works with multiple ODE solvers:

• Euler (basic)
• RF-Solver (second-order)
• FireFlow (velocity reuse, recommended)

Python API

from api import AdaEditPipeline

# Initialize
pipeline = AdaEditPipeline(
    model_path="checkpoints/flux-dev",
    device="cuda"
)

# Edit image
result = pipeline.edit(
    source_image="source.jpg",
    source_prompt="A photo of a cat",
    target_prompt="A photo of a dog",
    edit_object="cat",
    inject_schedule="sigmoid",
    use_channel_ls=True,
    num_steps=15,
    guidance=4.0,
    seed=42
)

# Save result
result.save("output.jpg")

Technical Details

Progressive Injection Weight Calculation

# Sigmoid schedule
w(t) = 1 / (1 + exp(k * (t/T_inj - 0.7)))

# Cosine schedule
w(t) = 0.5 * (1 + cos(π * t/T_inj))

# Linear schedule
w(t) = max(1 - t/T_inj, 0)

Channel Importance Estimation

# Calculate distribution difference for each channel
channel_diff = |mean(source_channel) - mean(random_channel)|

# Softmax normalization
channel_weight = softmax(channel_diff / temperature) * num_channels

Citation

If you use AdaEdit in your research, please cite our paper:

@article{li2026adaedit,
  title={AdaEdit: Adaptive Temporal and Channel Modulation for Flow-Based Image Editing},
  author={Li, Guandong and Chu, Zhaobin},
  journal={arXiv preprint arXiv:xxxx.xxxxx},
  year={2026}
}

Acknowledgments

• FLUX.1 - Base flow model
• FireFlow - Efficient ODE solver
• PIE-Bench - Evaluation benchmark

License

This project is licensed under the MIT License. See the LICENSE file for details.