Technical Summary: Flow of Truth (FoT)

1. Problem Statement

The rapid advancement of Image-to-Video (I2V) generation models (e.g., AnimateDiff, Wan2.2, Sora, Kling) enables the creation of highly realistic, temporally coherent videos from single images. While beneficial for creative production, this technology introduces significant risks of misinformation and content fabrication.

Existing forensic mechanisms face a critical limitation in this domain:

• Static vs. Dynamic: Traditional proactive forensics embed signatures into static images to localize tampering. However, I2V generation involves progressive visual evolution where pixels move, deform, disappear, or are semantically re-synthesized over time.
• Temporal Drift: In I2V, embedded forensic traces drift, deform, or become occluded as frames progress. Static spatial forensics, which assume fixed pixel alignment, become unreliable.
• The Core Challenge: The forensic task shifts from identifying where manipulation occurred in a single frame to tracing how protected evidence flows and transforms across time. Furthermore, defenders typically only have access to the attacker's released video frames; the original source image is unavailable for standard source-target motion estimation.

The paper identifies proactive temporal traceability as a new, unexplored forensic requirement: the ability to recover the original source image and trace evidence through motion and semantic changes without access to the original source during the forensic analysis phase.

2. Methodology: Flow of Truth (FoT)

The authors propose Flow of Truth (FoT), a proactive framework that reframes I2V generation from a pixel-motion perspective rather than simple frame synthesis. The core hypothesis is that if a forensic signal is designed to "follow" and "translate" with underlying pixel dynamics, it can remain synchronized with evolving target frames.

The framework operates through a four-stage pipeline:

A. Task Reformulation

The problem is defined as learning a mapping $\Phi: I_t \to \hat{I}_{t \to 0}$, where $I_t$ is an observed frame and $\hat{I}_{t \to 0}$ is the recovered source image. The final source is reconstructed by aggregating multi-frame predictions. This is reformulated as learning a dense displacement field (optical flow) from the source to the target, allowing for backward warping to recover the source.

B. Template Embedding

Instead of using fixed templates (sufficient for static localization but insufficient for motion tracking), FoT employs a learnable forensic template ($T$).

• An encoder $E$ embeds $T$ into the source image $I_0$ to create $I^T_0$.
• The embedding is optimized to be visually imperceptible (minimizing pixel distortion and LPIPS) while allocating energy to visually stable regions likely to maintain temporal consistency during synthesis.

C. Differentiable I2V Simulation

To avoid the computational prohibitions and bias of training directly on specific closed-source I2V models, FoT uses a differentiable simulation mechanism:

1. Compression/Reconstruction: A frozen Variational Autoencoder (VAE) simulates the compression and reconstruction dynamics of I2V models.
2. Motion Scattering: A motion field $M_i$ is randomly sampled from a bank of optical flow datasets (FlyingChairs, Sintel, Spring, etc.) to scatter pixels of the reconstructed image.
   This setup provides explicit ground-truth motion labels for training without requiring real I2V videos, ensuring the method generalizes across generators.

D. Motion Capture and Uncertainty Modeling

A template-guided flow prediction module estimates the dense displacement field between the embedded source and the simulated frame.

• Uncertainty Prediction: Recognizing that some regions (e.g., occluded areas or semantically re-synthesized content) are inherently unobservable, the model predicts a confidence map alongside the flow.
• Loss Function: A Mixture-of-Laplace loss is used to model both deterministic motion and uncertain cases (occlusion, new content), allowing the system to down-weight unreliable flow vectors during reconstruction.

E. Flow Reversal (Inference)

During inference, FoT decodes the surviving template evidence from the attacker's video frames.

1. It predicts the motion field from the source to the current frame.
2. It backward-warpes the frame to the source coordinate system.
3. It fuses multi-frame reconstructions using the predicted confidence maps as soft masks to generate the final recovered source image.

3. Key Contributions

1. Problem Identification: The paper defines proactive temporal traceability as a critical new forensic requirement for I2V, where evidence must be traced through motion and semantic drift.
2. Framework Design: FoT is the first proactive framework for source recovery in I2V. It uniquely models video generation as pixel-wise motion, learning a forensic template that co-evolves with pixel trajectories.
3. Generalization: The method is trained on a differentiable simulation of motion and compression, allowing it to generalize across both commercial (Kling, Dreamina) and open-source (CogVideoX, Wan2.2) I2V models without direct access to their weights.
4. Auxiliary Capabilities: The authors demonstrate that FoT acts as a general plug-in module that:
   • Enhances the robustness of copyright watermarks against geometric distortions and I2V transformations.
   • Improves tampering localization by leveraging motion-based features to detect content inconsistencies.

4. Experimental Results

The authors evaluated FoT on a custom benchmark comprising 204 source-prompt pairs across five motion scenarios (Face, Camera, Animal, Human-Environment, Multi-Human) and four I2V models.

• Truth Recovery: FoT significantly outperforms unprocessed "Forged" frames in recovering the original source image.
  • Frame-Level: FoT improved PSNR from ~20.4 to ~23.0 and SSIM from ~0.64 to ~0.76 on average across motion scales.
  • Video-Level: Even when 90% of the initial frames were discarded (a severe attack), FoT maintained a clear performance advantage over raw forged evidence, demonstrating resilience to temporal data loss.
• Motion Capture Accuracy:
  • Performance varied by model and scenario. CogVideoX yielded the most stable results (AEE ~6.88), while Dreamina S2.0 presented the highest challenges (AEE ~30.6).
  • Scenario Analysis: "Camera" motion (global, coherent) was easier to capture than "Human-Environment" or "Multi-Human" scenarios, which involve complex, non-rigid interactions.
  • Scale Analysis: Large motion magnitudes (>40px) consistently caused higher errors, echoing challenges in traditional optical flow, but FoT remained effective for structured large motions (e.g., camera zooms).
• Auxiliary Tasks:
  • Watermarking: FoT improved bit accuracy for geometric and I2V attacks when used as a post-embedding recovery layer for RoSteALS.
  • Tampering Localization: The framework achieved high Dice scores (>0.95 on clean inputs) in localizing edited regions, validating that motion-based features generalize to tampering detection.

5. Significance and Claims

The paper positions FoT as a foundational step toward proactive temporal forensics.

• Shift in Paradigm: It moves forensic analysis from static spatial localization to dynamic temporal tracing, addressing the specific nature of generative video where content evolves over time.
• Practical Utility: By recovering the "truth" (the source image) from manipulated videos, FoT provides a mechanism to verify authenticity even when the original source is lost or hidden by the attacker.
• Limitations and Future Work: The authors modestly acknowledge that FoT struggles with semantic re-synthesis (where the generator creates entirely new content not present in the source) and highly non-rigid, heterogeneous motions (e.g., complex human interactions). They suggest that future work should focus on high-level generative priors and region-aware reasoning to handle these "failure modes."

The work concludes that while tracking pixels through large, non-rigid displacements remains challenging, FoT establishes a viable pathway for tracing protected evidence through the temporal evolution of AI-generated video.