SwitchWorld FullFlow v2 adapters

SwitchWorld adds in-stream FP/TP viewpoint switching to LingBot-World. This repository contains the completed FullFlow v2 adapter sequence for both the high-noise and low-noise branches. The LingBot-World base checkpoint is required separately and is not duplicated here.

Files

File	Role
front_fp_high_rank128.pt	high-noise branch after FP expert training
front_tp_high_rank128.pt	high-noise branch after TP expert training
joint_high_rank128.pt	high-noise final adapter after transition training
front_fp_low_rank128.pt	low-noise branch after FP expert training
front_tp_low_rank128.pt	low-noise branch after TP expert training
joint_low_rank128.pt	low-noise final adapter after transition training

For normal inference, use joint_high_rank128.pt and joint_low_rank128.pt. The four intermediate checkpoints are included for reproducibility and ablations.

Download

hf download PencilHu/SwitchWorld --local-dir checkpoints/switchworld-fullflow-v2

Base model and code

• Base model: robbyant/lingbot-world-base-cam
• Code: yizhiqianbi/SwitchWorld
• Results: SwitchWorld Gallery

Training layout

Each noise branch was trained in three ordered stages. The FP expert is updated first, the TP expert is updated from that checkpoint, and the transition module is updated last while shared/FP/TP weights are frozen. Frozen-component digest checks passed for the completed local run.

Limitations

• These are adapters/checkpoints for the SwitchWorld research code, not a standalone Diffusers pipeline.
• Generation inherits the limitations and hardware requirements of LingBot-World and Wan2.2.
• The gallery includes research outputs from multiple stages, including baselines and failure cases; it should not be interpreted as a curated quality benchmark.
• Several evaluation metric interfaces in the codebase still have explicitly marked mock backends. Do not treat mock values as model claims.

License

Apache-2.0. The required upstream base model and datasets remain subject to their own licenses and terms.