The authors present WeLM-Image, a unified image foundation model that handles text-to-image generation, instruction-guided editing, and multi-reference synthesis in one interleaved sequence. It uses a Mixture-of-Transformers backbone with Mixture-of-Experts modules (31.7B stored parameters, about 5.4B active per target latent token), is built for single-GPU deployment in WeChat, and is distilled into a few-step Turbo variant with MeanFlow.
arXiv:2611.07392
The authors present SLA, a Spatial-Language-Action model initialized from a pretrained pi05 policy. During continual pre-training, it jointly optimizes action prediction and alignment between intermediate VLA visual embeddings and features from a frozen 3D perception model, using feature-dependent gates and probabilistic state dropout. All auxiliary components are used only during training, so the original inference architecture is kept.
arXiv:2611.08373
The authors release Open-HumanScale, an open recipe for scaling egocentric human video pretraining to 50,000 hours, built around a curated corpus (Open-HumanScale-50K) annotated with hand motion and language. They train HumanWAM-Vanilla, a world–action model pretrained only on human video, and HumanWAM-Goal, a visual-goal-conditioned variant trained with joint human–robot data.
arXiv:2611.06817
The authors propose TailSteer, a final-stage training framework that makes image editors more robust across repeated editing turns. It keeps early denoising dynamics close to the base model and learns preservation corrections near the trajectory tail, using the editor's own on-policy rollouts with unchanged source content as a reference.
arXiv:2611.06741