Lost in Edits? A $λ$-Compass for AIGC Provenance

TL;DR

This work tackles provenance attribution in the era of diffusion-based image synthesis and iterative text-guided editing. It introduces LambdaTracer, an alteration-free, inversion-based framework that uses a Box-Cox loss transform with adaptive λ selection to robustly distinguish model-generated images from edited ones without modifying generation pipelines. Through extensive experiments and ablations, LambdaTracer demonstrates superior attribution performance over a baseline, including in adversarial, iterative-edit scenarios, highlighting its practical value for copyright protection and content credibility in dynamic AI ecosystems. The approach offers a scalable, generalizable solution for tracing origins across diverse models and editing tools, with potential for improved interpretability and broader deployment in open settings.

Abstract

Recent advancements in diffusion models have driven the growth of text-guided image editing tools, enabling precise and iterative modifications of synthesized content. However, as these tools become increasingly accessible, they also introduce significant risks of misuse, emphasizing the critical need for robust attribution methods to ensure content authenticity and traceability. Despite the creative potential of such tools, they pose significant challenges for attribution, particularly in adversarial settings where edits can be layered to obscure an image's origins. We propose LambdaTracer, a novel latent-space attribution method that robustly identifies and differentiates authentic outputs from manipulated ones without requiring any modifications to generative or editing pipelines. By adaptively calibrating reconstruction losses, LambdaTracer remains effective across diverse iterative editing processes, whether automated through text-guided editing tools such as InstructPix2Pix and ControlNet or performed manually with editing software such as Adobe Photoshop. Extensive experiments reveal that our method consistently outperforms baseline approaches in distinguishing maliciously edited images, providing a practical solution to safeguard ownership, creativity, and credibility in the open, fast-evolving AI ecosystems.

Figures (6)

• Figure 1: An example of iterative text-guided editing InstructPix2Pix brooks2023instructpix2pix on model-generated images. The baseline LatentTracerwang2024trace struggles with cumulative perturbations, causing provenance inconsistencies. Our method, LambdaTracer, ensures robust tracing, effectively distinguishing authentic and manipulated content.
• Figure 2: Heatmap of PDF overlap in loss distributions across image categories. Rows and columns correspond to different image types: originally generated, single-edited, and iteratively edited images. P2P denotes the InstructPix2Pix brooks2023instructpix2pix, while CNet represents the ControlNet li2025controlnet. The suffix (e.g., P2P-1) indicates that the corresponding model was applied $n$ times iteratively. Each cell, ranging from $0$ to $1$, quantifies the degree of overlap, where higher values indicate greater similarity in loss distributions, and lower values represent better separability.
• Figure 3: The proposed pipeline consists of three steps: ① Image generation and edition by using stable diffusion models and text-guided editing methods respectively; ② Latent-space insertion and reconstruction loss evaluation via mean square error (MSE); ③ Data Transformation and Dynamic $\lambda$-selection strategies, including maximum likelihood, skewness minimization, and kurtosis minimization.
• Figure 4: Comparison of Baseline and LambdaTracer on manipulated images for different positive groups. Each group comprises various generative models, while the negative class covers various manipulations (e.g., Photoshop-based edits and iterative text-guided modifications). LambdaTracer, equipped with its $\lambda$-selection strategy, demonstrates consistent improvements over the Baseline
• Figure 5: Example of iterative text-guided editing on images generated by SD v2-base. The top row showcases images edited progressively using ControlNet, with iterations ranging from 1 to 5. The bottom row demonstrates images modified using InstructPix2Pix, also iterated from 1 to 5 times. This visualization highlights the cumulative effects of iterative modifications introduced by different editing models.