cosmos3_edge_image_HNPU

Cosmos3-Edge — Hexagon NPU (Image)

Prebuilt Qualcomm Hexagon NPU artifacts for the generation half of nvidia/Cosmos3-Edge — a diffusion model that runs on-device. It creates: given a text prompt, it synthesizes an image.

Capabilities

• Text-to-image — a diffusion-based generator that produces an image from a free-text prompt. Text in → image out.

This is a distinct network from the understanding model — a diffusion generator (with its own image decoder), not a language/vision-language model. It shares only the text-embedding table used to condition on the prompt.

Family — Cosmos3-Edge is shipped as two arch-pinned edge bundles by task:

• Cosmos3-Edge — Understanding — understanding (chat + vision), text output
• this repo — generation, image output

Artifacts are arch-pinned; this repo currently contains the v81/ directory.

For model architecture, intended use, limitations, and license terms, see the upstream model card: nvidia/Cosmos3-Edge. Distributed under the NVIDIA Open Model License inherited from the base model.