MagDiff: Multi-Alignment Diffusion for High-Fidelity Video Generation and Editing

TL;DR

MagDiff tackles the challenge of unified, high-fidelity video generation and editing by introducing a three-pronged alignment framework. It combines Subject-Driven Alignment to use subject-focused prompts, Adaptive Prompts Alignment to balance image-text control via learnable weights, and High-Fidelity Alignment to preserve subject details through a VAE-based pyramid encoder. Empirical results on UCF-101, MSR-VTT, DAVIS, and DreamBooth demonstrate superior or competitive performance across generation, editing, and fidelity metrics, with tuning-free inference. The approach promises practical impact for content creation by enabling precise subject control and robust editing within a single diffusion-based model, while acknowledging computational demands and potential limitations with challenging visual inputs.

Abstract

The diffusion model is widely leveraged for either video generation or video editing. As each field has its task-specific problems, it is difficult to merely develop a single diffusion for completing both tasks simultaneously. Video diffusion sorely relying on the text prompt can be adapted to unify the two tasks. However, it lacks a high capability of aligning heterogeneous modalities between text and image, leading to various misalignment problems. In this work, we are the first to propose a unified Multi-alignment Diffusion, dubbed as MagDiff, for both tasks of high-fidelity video generation and editing. The proposed MagDiff introduces three types of alignments, including subject-driven alignment, adaptive prompts alignment, and high-fidelity alignment. Particularly, the subject-driven alignment is put forward to trade off the image and text prompts, serving as a unified foundation generative model for both tasks. The adaptive prompts alignment is introduced to emphasize different strengths of homogeneous and heterogeneous alignments by assigning different values of weights to the image and the text prompts. The high-fidelity alignment is developed to further enhance the fidelity of both video generation and editing by taking the subject image as an additional model input. Experimental results on four benchmarks suggest that our method outperforms the previous method on each task.

Figures (14)

• Figure 1: Comparisons of our proposed unified diffusion model named (a) MagDiff, (c) MagDiff w/o HFA and other video diffusions, including (b) VideoCrafter1 chen2023videocrafter, (d) ModelScope wang2023modelscope, and (e) FateZero qi2023fatezero. The results show that our proposed MagDiff obtains the best visual performance (i.e. good text-and-image alignment and high fidelity) for both tasks of video generation and editing.
• Figure 2: An overview of our proposed Multi-alignment Diffusion (MagDiff), a unified diffusion method supporting both video generation and editing at the same time. Our MagDiff is comprised of three key components: 1) Subject-Driven Alignment (SDA) for unifying two tasks, 2) Adaptive Prompts Alignment (APA) for distinguishing the different controllability between homogeneous and heterogeneous modalities, and 3) High-Fidelity Alignment (HFA) for improving the quality of video generation or editing.
• Figure 3: The comparison between the image-driven method (VideoCrafter1 chen2023videocrafter) and our subject-driven method MagDiff. The subject-driven method can unify two tasks of video generation and editing but the image-driven method does not have this ability.
• Figure 4: Comparison of our Adaptive Prompts Alignment (APA) vs. Fixed Prompts Alignment. The APA uses two learnable parameters $\alpha_1$ and $\alpha_2$ to adaptively balance the trade-off of alignments between homogeneous and heterogeneous modalities.
• Figure 5: For qualitative evaluation, we compare our MagDiff with VideoCrafter1 chen2023videocrafter, I2VGen-XL zhang2023i2vgen-xl, and AnimateDiff (V3) guo2023animatediff on the generation task (orange dotted box) and compare with FateZero qi2023fatezero on the editing task (green dotted box).