// project
In ProgressCo-Developer
Robot World Video Decomposition Model
World ModelSimulationSAPIENIsaac Lab
Overview
- Evaluating single-image layer decomposition methods as the foundation for a robot-manipulation video layering pipeline feeding into a tri-view layered World Action Model (WAM).
- Compared four candidate approaches — See-through, LayerD, RevealLayer, and Qwen-Image-Layered — on domain alignment, semantic fixity, and occlusion recovery.
Progress
- Selected See-through (SDXL + consistency modules, up to 23 RGBA layers) as the primary candidate, with LayerD (BiRefNet + LaMa) as a lightweight sanity baseline.
- Ruled out Qwen-Image-Layered for poor performance on robot layers; deferred RevealLayer (FLUX.1-dev-based) pending further evaluation.
- Currently validating See-through on a curated set of 12–20 robot images, with Cosmos Predict 2.5 (2B) as the leading backbone candidate for the generation stage.