# Cosmos-Transfer2 Projects at AI Tinkerers

> Canonical HTML: https://aitinkerers.org/technologies/cosmos-transfer2
> Markdown URL: https://aitinkerers.org/technologies/cosmos-transfer2.md
> Technology record last updated: 2026-04-03T16:48:21Z
> Generated: 2026-09-23T00:40:02Z

A multi-controlnet world foundation model that transforms structured simulation data into physics-aware, photorealistic video for physical AI training.

Cosmos-Transfer2.5 is a diffusion transformer model engineered by NVIDIA to bridge the gap between simulation and reality for robotics and autonomous vehicles. By processing multimodal inputs like depth maps, segmentation masks, and RGB video, the 2B-parameter model generates high-fidelity world simulations that maintain strict temporal and physical consistency. It is 3.5 times smaller than its predecessor, Cosmos-Transfer1-7B, yet delivers superior prompt alignment and significantly lower error accumulation in long-horizon video generation. Developers use it to scale synthetic datasets with precise control over environmental variables (lighting, weather, and object placement) ensuring that perception models trained in Sim2Real pipelines translate reliably to real-world hardware.

- Official technology site: https://github.com/nvidia-cosmos/cosmos-transfer2.5
- Public AI Tinkerers demos and talks: 1
- Result page: 1 of 1

## Recent Public Talks and Demos

### [From Lab Walls to Real Fields: Fixing Data Scarcity with NVIDIA Cosmos-Transfer2.5 + FiftyOne (BioTrove Moth Edition)](https://raleigh.aitinkerers.org/talks/rsvp_ydT-2Uzjhhk)

A hands-on demo showing how to solve real-world data scarcity and domain shift using NVIDIA Cosmos-Transfer2.5 and FiftyOne. We begin by extracting rare moth samples from the BioTrove biodiversity dataset using BioCLIP embeddings, converting them into controllable video inputs, and generating edge-based control signals. NVIDIA Cosmos-Transfer2.5 then performs world-model–driven domain translation, transforming sterile lab images into realistic field environments. The outputs are validated and analyzed in FiftyOne using grouped datasets, embeddings, and semantic similarity search. This generative workflow yields 20–40% improvements on downstream classifiers by expanding rare classes and correcting dataset bias. This demo directly addresses three core challenges in real-world computer vision: (1) severe class imbalance and ecological domain shift, (2) absence of realistic field imagery for rare species, and (3) scalable validation of generative outputs. The session walks through the pipeline end-to-end: selecting scarce samples with BioCLIP, preparing video and edge-control inputs, running Cosmos-Transfer2.5 for domain transfer, and validating the generated outputs in FiftyOne using embeddings, grouped datasets, and semantic search.

- Event context: Raleigh AI Tinkerers: Second Meetup — 2025-12-10 — Raleigh
- Public talk page: https://raleigh.aitinkerers.org/talks/rsvp_ydT-2Uzjhhk

## Related Technologies

- [BioCLIP](https://aitinkerers.org/technologies/bioclip) ([Markdown](https://aitinkerers.org/technologies/bioclip.md)) — 1 public demo
- [BioTrove Dataset](https://aitinkerers.org/technologies/biotrove-dataset) ([Markdown](https://aitinkerers.org/technologies/biotrove-dataset.md)) — 1 public demo
- [FiftyOne](https://aitinkerers.org/technologies/fiftyone) ([Markdown](https://aitinkerers.org/technologies/fiftyone.md)) — 2 public demos
- [Real2Real](https://aitinkerers.org/technologies/real2real) ([Markdown](https://aitinkerers.org/technologies/real2real.md)) — 1 public demo