Amazon's DeltaWorld Beats NVIDIA Cosmos With 2,000x Fewer Computing Operations
DeltaWorld compresses each video frame into a single token, beating Cosmos with 35x fewer parameters and 2,000x fewer FLOPs while generating diverse futures in one pass
ProTakeaways
- One token per frame: DeltaTok encodes the VFM feature difference between consecutive frames into a single delta token, yielding a 1,024x token reduction at 512x512 resolution.
- Beats Cosmos at 2,000x lower cost: DeltaWorld-0.3B outperforms Cosmos-4B and Cosmos-12B on dense forecasting benchmarks using 2,000x fewer FLOPs and 35x fewer parameters.
- Diverse futures in one forward pass: Best-of-Many (BoM) training with K=256 Gaussian noise queries enables diverse future sampling without iterative diffusion.
- Delta prior prevents mode collapse: Predicting zero change defaults to copying the previous frame, stabilizing mean prediction quality to match discriminative baselines.
- Architecture-agnostic: Delta tokens also work in discriminative models (DINO-Foresight), matching performance with 2,048x fewer tokens and 5x less GPU memory.
- Open source: Code on GitHub and weights on Hugging Face; accepted as a CVPR 2026 Highlight.
Video world models , systems that predict plausible future states of a scene , are the backbone of autonomous driving, robotics, and planning agents. The problem is that they are brutally expensive. A single 512x512 frame typically requires 1,024 spatial tokens to represent, and generating multiple diverse futures means running many expensive forward passes through a large model. A new CVPR 2026 Highlight paper from Amazon and Eindhoven University of Technology asks a simple question: what if a frame only needed one token?
The answer is DeltaWorld, a generative world model built on a new tokenizer called DeltaTok. The system achieves better future predictions than NVIDIA's Cosmos (4B and 12B parameter models) while using over 35x fewer parameters and 2,000x fewer FLOPs , a gap so large it reshapes what's feasible for real-time deployment.
Why world models are stuck
There are two flavors of world model. Discriminative models take past frames and output a single deterministic prediction of the future. The problem: when the future is genuinely uncertain (a pedestrian might turn left or right), the model averages over possibilities and produces a blurry, unrealistic prediction that doesn't correspond to any real outcome. Generative models fix this by sampling multiple plausible futures, but they pay a steep computational price.
Three inefficiencies plague existing generative world models:
- They optimize for pixel-level fidelity, modeling irrelevant details like tree textures that don't affect downstream decisions.
- They require multiple sequential forward passes (e.g., iterative diffusion denoising) to produce even a single future sample.
- They represent each frame as a dense spatial grid of tokens, ignoring the fact that consecutive frames are mostly identical.
Recent work showed that predicting in the feature space of a Vision Foundation Model (VFM) , a large pretrained encoder like DINOv3 that produces rich semantic patch embeddings , rather than raw pixel space dramatically reduces the number of parameters needed. But those approaches remained discriminative. DeltaWorld is the first to combine VFM feature-space prediction with efficient generative sampling.
The core idea: encode the change, not the frame
The key insight is that consecutive video frames differ only in structured, low-dimensional ways. Backgrounds stay static; only a small portion of the scene changes between timesteps. Encoding the full scene at every step wastes capacity on redundant information.