Abstract
First-Frame Propagation (FFP) offers a promising paradigm for controllable video editing, but existing methods are hampered by a reliance on cumbersome run-time guidance. We identify the root cause of this limitation as the inadequacy of current training datasets, which are often too short, low-resolution, and lack the task diversity required to teach robust temporal priors. To address this foundational data gap, we first introduce FFP-300K, a new large-scale dataset comprising 300K high-fidelity video pairs at 720p resolution and 81 frames in length, constructed via a principled two-track pipeline for diverse local and global edits. Building on this dataset, we propose a novel framework designed for true guidance-free FFP that resolves the critical tension between maintaining first-frame appearance and preserving source video motion. Architecturally, we introduce Adaptive Spatio-Temporal RoPE (AST-RoPE), which dynamically remaps positional encodings to disentangle appearance and motion references. At the objective level, we employ a self-distillation strategy where an identity propagation task acts as a powerful regularizer, ensuring long-term temporal stability and preventing semantic drift. Comprehensive experiments on the EditVerseBench benchmark demonstrate that our method significantly outperforming existing academic and commercial models by receiving about 0.2 PickScore and 0.3 VLM score improvement against these competitors.
Introduction
Video editing is a complex task requiring "dual loyalty": the model must remain faithful to the user's edit (appearance) while preserving the original video's movement (motion). Current approaches are often split between instruction-based methods, which struggle with temporal consistency, and First-Frame Propagation (FFP). While promising, FFP models have historically been "brittle," requiring per-video fine-tuning or auxiliary inputs like depth maps. This paper argues that the primary bottleneck is a data gap—existing datasets are too short and low-resolution to foster robust temporal priors.
The FFP-300K Dataset
The centerpiece of this work is FFP-300K, a large-scale dataset containing approximately 290,441 original/edited video pairs. Unlike previous benchmarks, FFP-300K provides 720p resolution and 81-frame sequences. The dataset is built using a two-track pipeline:
- Local Editing Track: Employs VLMs and Grounded-SAM2 for precise object swaps and removals.
- Global Stylization Track: Uses motion-aware generative priors (VACE) and depth-guided conditioning to transform entire scenes into new artistic styles.
Methodology: FreeProp
The authors propose FreeProp, a framework designed for guidance-free propagation through two key innovations:
- Adaptive Spatio-Temporal RoPE (AST-RoPE): Standard positional embeddings are replaced with a dynamic system that uses a lightweight transformer to predict scaling factors based on the source video. This anchors appearance to the first frame while adjusting the temporal axis to match the video's specific motion.
- Self-Distillation with Identity Propagation: The model is trained to reconstruct the original source video from its first frame (Identity Propagation). This task serves as a "motion teacher," forcing the editing task to align its internal representations with ground-truth motion, thereby preventing semantic drift.
Results and Discussion
FreeProp sets a new state-of-the-art on the EditVerseBench benchmark, outperforming both academic models like EditVerse and commercial systems like Aleph.
Quantitative Performance
As illustrated in the radar charts, FFP-300K provides superior coverage in resolution, frame count, and diversity compared to existing datasets. Consequently, the FreeProp model achieves an improvement of approximately 0.2 in PickScore (perceptual quality) and 0.3 in VLM score (semantic correctness) over competitors. On temporal consistency metrics (DINO and CLIP), the model achieves scores near 0.991, indicating near-perfect stability.
Qualitative Analysis
Visual comparisons highlight FreeProp's ability to maintain high fidelity and consistent motion even across long 81-frame sequences. While competing models often suffer from flickering or lose the subject's identity, FreeProp remains stable across diverse tasks.
Conclusion
By scaling training data to 300K high-resolution pairs and introducing content-aware positional encodings, FreeProp demonstrates that robust temporal priors can be learned for generalizable video editing. This approach removes the need for slow, per-video fine-tuning, paving the way for more efficient and high-quality video-to-video foundation models.
Summary updated