Cosmos-Transfer2.5-2B
nvidia/Cosmos-Transfer2.5-2B · WFM
🤗 Model Card GitHub Paper Try the tool
WORLD-TO-WORLD · DIFFUSION TRANSFORMER

Physics-aware world generation
with adaptive multimodal control

Cosmos-Transfer2.5-2B is NVIDIA's diffusion-based world foundation model for Physical AI. Feed it a video plus a text prompt and it translates the structured control signals inside that video — depth, edges, segmentation, blur — into a new, photorealistic world state. Built for Sim2Real augmentation of autonomous-vehicle and robotics data.

2.36B
Parameters
93 frames
≈5 s · 720p @ 16 fps
4 controls
+ multi-control & multiview
65.4 GB
Inference VRAM

01How a transfer works

World-to-world: an existing video's structure is preserved, its appearance is regenerated.

🎥
Input video
Any RGB clip — robot arm, dashcam, drone. Multiples of 93 frames work best.
──▶
🎛️
Control branches
Depth / edge / seg / blur signals extracted per frame, weighted per modality.
──▶
🌍
New world state
Same motion & layout, entirely new appearance steered by the text prompt.

02Control modalities

Official checkpoints on the Hub — click one to load it into the spec generator below.

03Inference spec generator

Compose a controlnet_spec exactly as the official examples/inference.py consumes it, then copy the one-liner to run it on your own GPU.

Model card guidance: under 300 words; describe scene, key objects, background and motion within the 5-second window.
Distilled model: exactly 93 frames, --model=edge/distilled.
Why no in-browser generation? This model needs 65.4 GB of GPU VRAM (above any single Space GPU), a license-gated checkpoint, and ~2–15 min of diffusion per 93-frame clip. The official path is local / cluster inference from a cloned repo — this tool gives you a ready-to-run spec instead of a fake progress bar.
controlnet_spec.json 

          

            run it
            
          

          

        

      

    

  


  

    
04Quickstart

    
From the official repo — the only supported runtime for the full transfer pipeline.

    


# clone & setup (CUDA 12.x, Ampere/Hopper/Blackwell)
git clone https://github.com/nvidia-cosmos/cosmos-transfer2.5.git
cd cosmos-transfer2.5
uv sync --extra cu128
python -m cosmos_transfer2.utils.dotdict >/dev/null  # sanity import
# accept the NVIDIA Open Model License on the Hub, then:
huggingface-cli login

# run your generated spec (see panel above)
python examples/inference.py -i my_spec.json -o outputs/

      
Diffusers note: the snippet on the model card is auto-generated boilerplate — the Hub repo ships raw .pt checkpoints (general/<modality>/*_ema_bf16.pt), not a diffusers layout. Use the cosmos-transfer2.5 repo for real inference.

    

  


  

    
05Model lineage

    

      

        
2025-10-06
Cosmos-Transfer2.5 release
Multi-control transfer models + Cosmos-Predict2.5, open on GitHub and the Hub.

        
2025-11-07
Autoregressive long videos & multiview
Sliding-window generation for longer clips; new multiview cross-attention module; Blackwell support.

        
2025-12-12
Updated checkpoints
Refreshed blur / depth / seg / edge weights; on-the-fly control computation; single-view post-training recipes.

        
2026-02-23
Distilled edge checkpoint
~7.5× faster diffusion (4 sampling steps) for short, strictly-93-frame edge transfers.

        
2026-06-01
Cosmos 3
Successor platform unifying language, video, audio and actions — Transfer2.5 is now in maintenance mode.

      

    

  


  

    
06Performance & requirements

    

      

        
GPU93-frame gen (seg control)E2E (121-frame input)

        
          
B20092 s187 s

          
H100 PCIe264 s534 s

          
H100 NVL446 s895 s

          
H20684 s1370 s

        
      

      
720p, 16 fps, guardrails off — official numbers from the repo's inference guide. Single-GPU VRAM requirement: 65.4 GB.

    

    

      
Temporal inconsistencyLong or high-resolution generations can drift; motion instability is a known limitation.

      
Control adherenceThe model may not follow the control video precisely, especially with multiple weighted controls.

      
Prompt scopeText conditions the whole 5-second window; <300 words recommended.

      
Content riskUnconditional video generation — output is not filtered for all regions/use cases by default (guardrails available in the repo).