TCAN
Using a diffusion model to achieve temporal consistency in portrait animation
- Appearance Attitude Adaptation (APPA layer): Maintain attitude information from the frozen control network while preserving the appearance of the source image.
- Time controlled network: prevents generated videos from crashing due to sudden and incorrect posture changes.
- Attitude driven temperature map: Reduce flicker in static areas by smoothing the attention score of the time layer during the inference phase.
- Time consistency: Ensure the coherence of character postures during the animation process.
- Generalization ability: able to adapt to animation generation in different fields and identities.
- Background preservation: Maintain consistency of the source image background during the animation process.
- Multi identity animation: capable of transferring actions to characters of different identities or animated characters.
Product Details
TCAN is a novel portrait animation framework based on diffusion models, which can maintain temporal consistency and generalize well to unseen domains. This framework ensures that the generated video maintains the appearance of the source image while following the posture of the driving video, while maintaining consistency in the background, through unique modules such as the Appearance Attitude Adaptive Layer (APPA layer), Time Control Network, and Attitude Driven Temperature Map.
