FreqEdit: Preserving High-Frequency Features for Robust Multi-Turn Image Editing

FreqEdit is a training-free framework that mitigates quality degradation in multi-turn image editing by preserving high-frequency information through feature injection, adaptive spatial modulation, and path compensation, thereby achieving superior identity preservation and instruction following across numerous consecutive iterations.

The Gist

Imagine you are a digital artist using a magical paintbrush to edit a photo. You want to change the background, then add a hat, then change the shirt color, then make the person smile.

The Problem: The "Blurry Copy" Effect
With current AI tools, the first edit looks great. But if you keep editing the result of the previous edit (like a game of "telephone"), the image starts to rot.

• The Face: The person's face starts to melt or look like a different person.
• The Skin: The skin turns into a smooth, plastic mannequin because all the tiny pores and texture vanish.
• The Edges: The outlines of objects get jagged and weirdly sharp.

The paper calls this "High-Frequency Loss." Think of an image like a song. The "low frequencies" are the bass and melody (the big shapes, the colors, the general idea). The "high frequencies" are the high-pitched cymbals and crisp details (skin pores, hair strands, fabric texture).

Current AI tools are like a bad microphone that only picks up the bass. Every time you edit, it loses a little more of the "cymbals." After 10 edits, the song is just a dull thud.

The Solution: FreqEdit (The "Detail Preserver")
The authors created a new tool called FreqEdit. It doesn't need to be retrained; it just works on top of existing AI models. It solves the problem with three clever tricks, using a metaphor of restoring a faded photocopy:

1. The "Reference Blueprint" (High-Frequency Injection)

Imagine you have a pristine, high-quality original photo (the "Context Image"). Every time you make a new edit, FreqEdit looks at that original photo and says, "Hey, look at how sharp the eyelashes are in the original. Let's steal that sharpness and paste it onto our new edit."

• How it works: It uses a mathematical tool called a Wavelet Transform (think of it as a "frequency filter") to separate the "boring big shapes" from the "exciting tiny details." It takes the tiny details from the original and injects them into the new image before the AI finishes drawing it.

2. The "Smart Mask" (Adaptive Injection)

Here's the tricky part: You don't want to steal details from the original if you are trying to change something.

• Scenario: You want to change a red car to a blue truck.
• The Mistake: If you just paste the original details everywhere, the AI might try to keep the red car's shape and the blue truck's shape, resulting in a weird red-blue monster.
• The Fix: FreqEdit is smart. It looks at where you are editing.
  • If you are changing the background: It says, "Okay, I'll steal the details for the background to keep it crisp."
  • If you are changing the car: It says, "Stop! Don't steal the car's details, or the new blue truck won't form. Let the AI create new details for the truck."
  • It creates a dynamic map that decides exactly how much "detail stealing" to do in every square inch of the image.

3. The "Course Correction" (Path Compensation)

Sometimes, if you force too many details back into the image, the AI gets confused. It might start drawing a "ghost" version of the old object and the new object at the same time (like a double-exposure photo).

• The Fix: FreqEdit acts like a GPS. Every few steps of the drawing process, it checks: "Are we still heading toward the user's instruction?" If the AI has drifted too far because of all the detail injection, FreqEdit gently nudges the drawing back onto the correct path, ensuring the final result actually looks like what you asked for.

The Result

Without FreqEdit, a 10-step editing session looks like a blurry, melted mess.
With FreqEdit, you can edit the image 10, 15, or even 20 times, and the person still looks like the same person, the skin still has texture, and the edges stay crisp.

In Summary:
FreqEdit is like a digital bodyguard for your photos. While the AI tries to simplify the image (which causes the "melting"), FreqEdit constantly whispers, "No, keep the details! Remember the original texture! Don't let the face melt!" ensuring your photo stays sharp and true to life, no matter how many times you edit it.

Technical Summary

1. Problem Statement

Instruction-based image editing using natural language has become a powerful tool for visual manipulation. While recent models (e.g., FLUX.1 Kontext, Qwen-Image) excel at single-turn edits, they suffer from severe quality degradation when subjected to multi-turn (iterative) editing.

Through systematic analysis, the authors identify three primary failure modes that emerge after approximately 5–10 consecutive editing turns:

1. Subject Deformation: The geometric structure and identity of the main subject progressively deviate from the original.
2. Edge Over-sharpening: Boundaries become artificially enhanced and unnatural.
3. Texture Collapse: Fine-grained details (e.g., skin pores, fabric textures) degrade into smooth surfaces or artifacts.

Root Cause: The authors hypothesize and validate that this degradation is caused by the progressive loss of high-frequency (HF) information across iterations. As the denoising process repeats, the model increasingly relies on its learned priors (canonical training data representations) rather than the specific details of the input image, causing the image to regress toward an "average" appearance.

2. Methodology: FreqEdit

FreqEdit is a training-free framework designed to stabilize multi-turn editing by selectively reinforcing high-frequency features. It consists of three synergistic components:

A. Wavelet-Based Feature Injection

Instead of injecting raw pixel data, FreqEdit operates in the frequency domain using a 2-level Discrete Wavelet Transform (DWT).

• Reference Velocity Construction: For the current editing turn $k$, a "reference velocity field" ($v_{ref}$) is constructed from the context image (the input for the current turn). This velocity represents the trajectory required to reach the context image from the current noisy state.
• Decomposition: Both the predicted editing velocity ($v_{edit}$) and the reference velocity ($v_{ref}$) are decomposed into low-frequency (LL) and high-frequency (D) components.
• Injection: Only the high-frequency components ($D_{ref}$) from the reference velocity are injected into the editing velocity. The low-frequency components (global structure/semantics) are left untouched to ensure the editing instruction is followed.
• Formula: The corrected high-frequency coefficients are computed as:
  $$\tilde{D}^{(\ell)} = D_{edit}^{(\ell)} + \alpha (D_{ref}^{(\ell)} - D_{edit}^{(\ell)})$$
  where $\alpha$ controls the injection strength.

B. Adaptive Injection Strategy

Uniform injection can suppress desired semantic changes in regions being edited. FreqEdit introduces a spatially-adaptive mechanism:

• Divergence Map: It calculates the L2 norm difference between $v_{edit}$ and $v_{ref}$ to identify regions undergoing semantic change (high divergence) versus regions that should remain stable (low divergence).
• Modulation: The injection strength $\alpha$ is modulated spatially. Regions with low divergence (stable areas) receive strong HF injection to preserve details. Regions with high divergence (editing areas) receive attenuated injection to allow the model to execute the transformation without conflict.

C. Path Compensation Mechanism

Continuous HF injection can sometimes lead to "ghosting artifacts" (where conflicting visual elements from the reference and editing velocities appear simultaneously) or over-constrain the generation.

• Trajectory Realignment: The method periodically (every $n$ steps) calculates the cumulative velocity difference ($\Delta v$) between the corrected trajectory and the pure editing trajectory.
• Correction: It applies a compensation step to the latent state, effectively "steering" the denoising trajectory back toward the intended editing direction while retaining the injected high-frequency details. This ensures the final output aligns with the instruction while maintaining fidelity.

D. Quality Guidance (Noise Suppression)

For models prone to noise accumulation (like FLUX.1 Kontext), a final refinement step blends the editing velocity with an auxiliary velocity derived from the original source image during the final denoising steps. This suppresses accumulated noise artifacts without altering the semantic content.

3. Key Contributions

1. Diagnosis of Multi-Turn Degradation: The paper provides empirical evidence that the loss of high-frequency features is the fundamental cause of identity drift and texture collapse in iterative editing.
2. Training-Free Framework: FreqEdit requires no model retraining or fine-tuning. It operates as a plug-in module for existing rectified flow or diffusion models.
3. Synergistic Components: The combination of wavelet-based injection, spatial adaptivity, and path compensation addresses the trade-off between preserving identity and following complex instructions.
4. State-of-the-Art Performance: It significantly outperforms existing baselines in maintaining consistency over 10+ turns.

4. Experimental Results

The authors evaluated FreqEdit on FLUX.1 Kontext and Qwen-Image against seven state-of-the-art baselines (including MTC, VINCIE, Nano Banana, and Seedream 4.0) using a dataset of 70 images with 10 sequential editing turns each.

• Qualitative Results: Visual comparisons show that baseline models suffer from severe body deformation and texture loss by Turn 7–10. FreqEdit maintains coherent subject identity, sharp edges, and natural textures throughout the entire sequence.
• Quantitative Metrics:
  • Consistency: FreqEdit achieves the highest scores in subject and background consistency (measured by DINOv2 similarity and L1 distance).
  • Instruction Following: While maintaining high consistency, FreqEdit retains competitive instruction-following scores (only a marginal drop compared to base models, e.g., 0.803 $\to$ 0.790 for FLUX.1).
  • Perceptual Quality: It significantly outperforms baselines in LPIPS and human preference scores.
• Human Preference: In a user study, Qwen-Image + FreqEdit achieved the highest preference score, ranking above the proprietary "Nano Banana" model and all other baselines.

5. Significance

• Enabling Complex Workflows: FreqEdit makes multi-turn editing practically viable for professional workflows (e.g., photographers refining portraits through dozens of steps), a capability previously limited by rapid quality decay.
• Theoretical Insight: It shifts the focus of editing stability from semantic alignment to frequency preservation, offering a new perspective on how generative models degrade over time.
• Accessibility: As a training-free solution, it can be immediately applied to improve the robustness of any current or future instruction-based image editing model without the computational cost of retraining.

In conclusion, FreqEdit solves the "iterative editing curse" by treating high-frequency details as critical identity anchors, using a sophisticated injection and compensation strategy to preserve them across multiple editing turns.