apiVersion: capsule.dev/v0.1 kind: Capsule name: cverse-avatar-inference-adapter version: 0.1.0 type: adapter purpose: summary: 'Provides adapters for real-time avatar generation models, specifically FlashHead and SoulX-LiveAct. This capsule integrates these complex model implementations into the core inference service. ' owns: - Base avatar plugin interface - FlashHead model implementation and utilities - SoulX-LiveAct model implementation and utilities does_not_own: - The core inference service - Other AI model types (ASR, LLM, TTS, etc.) interfaces: provides: - kind: library name: avatar-plugin description: Avatar model implementations conforming to the inference core's plugin interface. dependencies: capsules: - name: cverse-inference-core version: '>=0.1.0' agent: summary_for_ai: 'An AI agent working on this capsule would focus on improving avatar generation quality, optimizing model performance (e.g., latency, VRAM usage), or integrating new avatar models. It requires deep knowledge of real-time video generation, computer vision, and machine learning frameworks (e.g., PyTorch). ' avoid: - Modifying the core inference service or other AI model types. verification: invariants: - Avatar plugins must generate video streams from audio/visual inputs. - Plugins must adhere to the defined avatar plugin interface. - Output video must be synchronized with audio. x-reuse: notes: 'This capsule bundles two large, complex model implementations (FlashHead and SoulX-LiveAct). While the plugin interface is generic, the models themselves are highly specialized. Consumers would likely replace these with their own avatar models or use pre-trained versions. Model weights and specific training data are not included but are implied dependencies for these models to function. ' x-reconstruct: install: install.json