Instructions to use nvidia/omni-dreams-models with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- Cosmos
How to use nvidia/omni-dreams-models with Cosmos:
# No code snippets available yet for this library. # To use this model, check the repository files and the library's documentation. # Want to help? PRs adding snippets are welcome at: # https://github.com/huggingface/huggingface.js
- Notebooks
- Google Colab
- Kaggle
Enable HF download tracking via library_name: cosmos
What this does
Enables Hugging Face download-stats tracking for this repo by adding metadata to the model card. Built on NVIDIA Cosmos.
Change
README.md frontmatter only:
library_name: cosmosβ enables download counting (see below)pipeline_tag: image-to-video- license metadata (NVIDIA Open Model License)
- discovery tags (
cosmos,world-model,autonomous-driving,diffusion-transformer,image-to-video)
The model-card body is unchanged (keeps the "Built on NVIDIA Cosmos." attribution).
Why this enables tracking
Per the Hub download-stats docs, a repo's download count = HTTP requests (GET/HEAD) to its query files, and the query-file set is chosen by the repo's library_name.
Today this repo declares no library_name, so downloads are not counted at all. Note a bare config.json would not fix this: the default filter only counts config.json itself, and these checkpoints load via torch.load(*.pt) β nobody fetches config.json, so the count would stay ~0.
cosmos is already registered in huggingface.js with:
countDownloads: path:"config.json" OR path_extension:"pt"
The path_extension:"pt" clause means every .pt request is counted, with no extra files needed. This covers all four published checkpoints:
single_view/2b_res720p_30fps_i2v_hdmap_distilled.pt(released inference checkpoint)single_view/distilled/β¦_model.pt(L0 distilled)single_view/teacher/β¦_model.pt(L1b teacher)single_view/student-init/β¦_model.pt(L2a student-init)
cosmos is also accurate: these are Cosmos-architecture diffusion transformers fine-tuned from nvidia/Cosmos-Predict2.5-2B. The cosmos entry defines no usage snippet, so no misleading code widget is injected β the only page change is a small "Cosmos" library pill plus inclusion in the library=cosmos filter.
Notes
- The "Cosmos" library pill has no per-repo editable description (it is defined globally in huggingface.js); the OML-required "Built on NVIDIA Cosmos." attribution remains in the model-card body.
- Alternative considered: a FlashDreams-branded pill would require first registering
flashdreamsin huggingface.js withcountDownloads: path_extension:"pt", and tracking would only start once that PR merges. This PR usescosmosso tracking works immediately, andlibrary_namecan be switched later.