Contextualized Diffusion Models for Text-Guided Image and Video Generation
This work addresses a key bottleneck in text-guided visual generation for AI and creative applications, offering an incremental improvement over existing diffusion models.
The paper tackles the inconsistency between forward and reverse processes in text-guided visual diffusion models, which limits semantic alignment, by proposing ContextDiff, a contextualized diffusion model that incorporates cross-modal context into both processes. It achieves state-of-the-art performance in text-to-image generation and text-to-video editing, significantly enhancing semantic alignment as shown in evaluations.
Conditional diffusion models have exhibited superior performance in high-fidelity text-guided visual generation and editing. Nevertheless, prevailing text-guided visual diffusion models primarily focus on incorporating text-visual relationships exclusively into the reverse process, often disregarding their relevance in the forward process. This inconsistency between forward and reverse processes may limit the precise conveyance of textual semantics in visual synthesis results. To address this issue, we propose a novel and general contextualized diffusion model (ContextDiff) by incorporating the cross-modal context encompassing interactions and alignments between text condition and visual sample into forward and reverse processes. We propagate this context to all timesteps in the two processes to adapt their trajectories, thereby facilitating cross-modal conditional modeling. We generalize our contextualized diffusion to both DDPMs and DDIMs with theoretical derivations, and demonstrate the effectiveness of our model in evaluations with two challenging tasks: text-to-image generation, and text-to-video editing. In each task, our ContextDiff achieves new state-of-the-art performance, significantly enhancing the semantic alignment between text condition and generated samples, as evidenced by quantitative and qualitative evaluations. Our code is available at https://github.com/YangLing0818/ContextDiff