Customized Appearance And Motion Composition

[ECCV 2024 Oral] MotionDirector: Motion Customization of Text-to-Video Diffusion Models.

Unique: Implements LoRA weight composition in the attention module where spatial and temporal LoRAs are applied to different attention heads/layers without interference, enabling true orthogonal composition rather than simple weight addition.

vs others: Provides finer control than single-LoRA approaches and avoids retraining for each character-motion combination, unlike traditional animation pipelines requiring separate motion capture per character.

via “motion intensity and style control”

LivePortrait — AI demo on HuggingFace

Unique: Implements motion intensity control through learned scaling functions that preserve anatomical plausibility across intensity ranges, rather than naive vector scaling which produces distortion at extreme values, by constraining scaled landmarks to valid facial geometry

vs others: More intuitive than low-level parameter adjustment and more flexible than fixed presets because it provides continuous control with automatic constraint satisfaction, enabling users to explore the full expression space without manual tuning

via “animation-style-customization”

via “character and asset customization”