WorldFoundry
English中文

Cosmos Transfer 2.5

Method overview: Cosmos Predict 2.5’s world-model architecture from the paper.

NVIDIA sim-to-real arm: control video in, photoreal world video out. Native VACE on the Predict 2.5 runner; 2B edge is the recorded route.

Edit this pageGitHubMarkdown

Cosmos Transfer 2.5 is the sim-to-real / world-to-world arm of the Cosmos 2.5 generation. A control video (raw RGB, edge map, or simulation render) plus a prompt go in; a photoreal world video that follows that control comes out. This card is not Predict 2, not Predict 2.5, and not Cosmos 3. Predict cards synthesize future world video from text/image/video. Transfer translates an already-structured control signal.

Method overview: Cosmos Predict 2.5’s world-model architecture from the paper.
Method overview: Cosmos Predict 2.5’s world-model architecture from the paper.

It reuses Predict 2.5's native stack — Reason1 conditioner, flow-UniPC, shared Wan VAE, same loader — and adds only the checkpoint-shaped VACE control branch (four blocks at base layers 0/7/14/21, plus an 8-modality control embedder on the 28-layer base). No Diffusers, no ControlNet backend, no NVIDIA runtime package. Paper: World Simulation with Video Foundation Models for Physical AI (the same 2.5 write-up that covers Predict 2.5). Code: nvidia-cosmos/cosmos-transfer2.5.

The recorded WorldFoundry variant is cosmos-transfer-2.5-2b only. Catalog notes a checkpoint-backed 2B edge run at 1280×704, 93 frames, 35 steps. A 14B Hugging Face card exists (nvidia/Cosmos-Transfer2.5-14B) but is access-gated and not a verified WorldFoundry route here. Weights use the gated nvidia-open-model-license.

Build your run

Pipeline
CosmosTransfer2p5Pipelineworldfoundry.pipelines.cosmos.pipeline_cosmos_transfer2p5
TaskDefault default
Environmentworldfoundry-unified-cu128
DeviceCUDA 12.8
Runnerworldfoundry.pipeline
Loadingfrom_pretrained
Invocationnative_or_standard_invocation
worldfoundry-eval run \
  cosmos-transfer-2.5-2b \
  --pipeline.task-profile default \
  --json

What it is for

Turn structured control into photoreal world video. Launch worldfoundry-eval run cosmos-transfer-2.5-2b. Required inputs: prompt, image, video. The actions field is control_video. Artifact: cosmos-transfer2.5-2b-controlled-video.mp4.

Give the native route a raw RGB control clip and it extracts Canny edges itself. Set control_is_preprocessed only when the clip is already an edge map. This is not text-to-world (Predict 2.5) and not T5-conditioned video2world (Predict 2).

Score controlled world video on WorldModelBench and Physics-IQ. Those are catalog entry points, not ingested WorldFoundry scores.

Running in WorldFoundry

CosmosTransfer2p5Pipeline runs in worldfoundry-unified-cu128 (Python 3.11, CUDA 12.8). Stage the official 2B edge checkpoint nvidia/Cosmos-Transfer2.5-2B and the Predict 2.5 VAE plus Reason1 text encoder. If those gated files are missing, the runtime profile still marks exact checkpoint-key / GPU parity as pending even though the catalog records the 2B edge geometry above.