My Contribution — World Navigation

Core contributor on the World Navigation capability — data processing and training validation. The model predicts next first-person / third-person views from discrete action text inputs, with physical behaviors (gravity, collision, buoyancy, momentum) emerging from multimodal training rather than explicit simulation.

Method

ConsistencySolver is a lightweight, learnable high-order ODE solver derived from general linear multistep methods, optimized via Reinforcement Learning (PPO). It adapts integration coefficients to each sampling step, improving both preview fidelity and consistency with the full-step output—without modifying the base diffusion model.

• Diffusion Preview Paradigm: Fast few-step preview → user evaluation → full-step refinement only when satisfied
• Learnable ODE Solver: Adaptive multistep coefficients predicted by a lightweight MLP, trained end-to-end via RL
• Preview Consistency: Preserves the deterministic PF-ODE mapping so preview closely matches the final output
• FID on-par with 47% fewer steps vs. Multistep DPM-Solver; outperforms distillation baselines

Method

PromptRL jointly trains language models and flow-matching models in a unified RL loop for text-to-image generation, using LMs as adaptive prompt refiners.

• Unified RL Training: Joint optimization of both LM and FM in a single loop
• Exploration Collapse Solution: Overcomes insufficient diversity through adaptive prompting
• Test-time Adaptation: Adaptive prompt refinement at inference