Loading…
From a small amount of real data to diverse training data: lighting transfer, object removal/replacement, EGO-view synthesis and pixel-level auto annotation.
Multi-light environment transfer (stage light, window light) with physical consistency.
Precise removal of objects (eraser, apple, bottle cap); apple-to-lemon, Rubik's-cube-to-sponge replacement.
Add or remove objects in EGO view — peelers and plastic bags removed while preserving physics; phones and drinks naturally blended into scenes.
Hand-pose parsing and precise reconstruction for manipulation-level synthesis.
Automatic generation of instance segmentation, hand pose, depth, optical flow and object trajectory annotations.
High-speed physics-model inference quickly generates high-fidelity synthetic samples from real data.
Datasets captured with in-house devices and released open-source for embodied-AI / VLA training — real-world scenarios, multimodal and time-aligned, continuously updated.