PrimeSim

PrimeSim is an action-conditioned multi-view world model based on Wan2.2-TI2V-5B: it takes the first frame from each camera view, a text prompt, and an action trajectory as input to generate subsequent frames. The authors state that it was trained on 30,110 hours of data covering 22 robot platforms, leading in four out of five metrics on the WorldSimProbe simulation benchmark and achieving the best overall score on EWMBench. It provides base weights from mid-training (for fine-tuning) and causal few-step inference weights from stage four; during inference, three views are generated simultaneously at 192×256, with the authors claiming performance exceeding 10 FPS.

PrimeSim is an action-conditioned multi-view world model based on Wan2.2-TI2V-5B: it takes the first frame from each camera view, a text prompt, and an action trajectory as input to generate subsequent frames. The authors state that it was trained on 30,110 hours of data covering 22 robot platforms, leading in four out of five metrics on the WorldSimProbe simulation benchmark and achieving the best overall score on EWMBench. It provides base weights from mid-training (for fine-tuning) and causal few-step inference weights from stage four; during inference, three views are generated simultaneously at 192×256, with the authors claiming performance exceeding 10 FPS.

Why it matters Using generative video instead of physical simulation to preview robot operations, generating three camera views together, with commercially usable weights.

Hugging Face Likes 2 · Hugging Face Trending Score 2 · Hugging Face Downloads 201

License: Commercial use allowed · ComfyUI: Not available · VRAM: 24 GB

Reasons

  • Apache-2.0 license; the paper, project page, inference code, and training code are all publicly available

Usage

  • Download the text encoder, VAE, and tokenizer from Wan2.2-TI2V-5B, configure paths according to the GitHub repository, then run examples using primesim_posttrain_g2

Risks

  • Recommended A100 or H100 80GB; the 5B backbone and 11 GB text encoder must reside in memory simultaneously, which cannot fit on a 24 GB GPU
  • Training data is not public; evaluation numbers are self-reported by the authors

Open

抱脸 world-model 标签

Explore the latest AI Radar updates