NVIDIA Cosmos Transfer

NVIDIA's controllable video-generation model that produces photorealistic video conditioned on structure maps like depth, segmentation, and edges.

Cosmos Transfer is the conditional-generation branch of NVIDIA's Cosmos family, using a multi-branch ControlNet structure (feeding extra structure maps in as generation conditions). Given spatial control signals such as a segmentation map, depth map, edge map, blurred footage, high-definition map, or lidar, it outputs photorealistic video that keeps the same layout and motion, but can change lighting, weather, and materials. Cosmos-Transfer1, from March 2025, can assign different weights to different control signals at different locations in the frame; Cosmos-Transfer2.5, from October 2025, has 2B parameters and is built on Predict2.5, with a low-latency distilled version following in February 2026. Its main use is data augmentation: turning simulator-rendered robot or driving footage into photorealistic video to narrow the sim-to-real gap, or turning a piece of real footage into rainy or nighttime versions to expand training data.

ExampleRender a robot-arm grasping video in Isaac Sim, extract its depth and segmentation maps, and feed them to Cosmos Transfer; swapping in different text prompts yields multiple photorealistic versions with different backgrounds and lighting, used to train a policy more robust to visual distractions.

Also called

Cosmos-Transfer1, Cosmos-Transfer2.5

Related

NVIDIA Cosmos · Sim-to-Real Transfer · Sim-to-Real Gap (Reality Gap) · Generative Data Augmentation · Synthetic Data · NVIDIA Cosmos Predict

Sources

Cosmos-Transfer1: Conditional World Generation with Adaptive Multimodal Control (arXiv 2503.14492)
nvidia-cosmos/cosmos-transfer2.5 GitHub 仓库 (Chinese)

As of

2026-02