| # Repository layout |
|
|
| Paths identify artifact type, model family, and a coherent release or method group. |
| The catalog records tasks and modalities; folders do not impose a t2v/i2v pipeline. |
| This supports diffusion and non-diffusion approaches, image and video methods, full models, |
| adapters, face replacement, reenactment, avatar pipelines, and future artifact types. |
|
|
| ```text |
| models/<family>/<group>/<artifact-type>/ |
| workflows/<runtime>/<family>/<group>/ |
| scripts/<family>/<group>/ |
| integrations/<runtime>/ |
| reports/<family>/<group>/ |
| experiments/<study>/ |
| manifests/artifacts.json |
| docs/ |
| licenses/ |
| ``` |
|
|
| | Path | Contents | |
| | --- | --- | |
| | `models/` | Weights, model cards, licenses, and notices. Artifact types can be `loras`, `checkpoints`, `encoders`, `vaes`, `onnx`, or another descriptive directory. | |
| | `workflows/` | Runnable workflows organized by runtime. Store GUI JSON at the group root and API JSON in `api/`. Other runtimes use their own directories. | |
| | `scripts/` | Extraction, conversion, training, inference, and evaluation scripts with explicit arguments and a usage document. | |
| | `integrations/` | Runtime-specific loaders, nodes, or plugins shared by compatible workflows. | |
| | `reports/` | Numerical extraction and validation results, without test media or execution-history dumps. | |
| | `experiments/` | Completed studies: reports, configurations, numerical results, runtime evidence, source/published SHA provenance, model URLs and reproducibility instructions. Exclude media, weights, embeddings and raw histories. | |
| | `manifests/` | Machine-readable artifact inventory and compatibility information. | |
| | `docs/` | Collection-wide conventions and instructions. | |
| | `licenses/` | Shared license texts; model-specific licenses also accompany model groups. | |
|
|
| ## Adding artifacts |
|
|
| 1. Choose lowercase kebab-case family and group names, such as `minimax-h3/viggle-delta`. |
| 2. Add weights under their artifact type. Include a group README, license, provenance, and notices. |
| 3. Place companion workflows, scripts, integrations, and reports under their corresponding roots. |
| 4. Add one catalog entry per independently selectable model variant. Record paths, roles, bytes, |
| SHA-256, kind, tasks, modalities, exact base compatibility, dependencies, license, and validation. |
| 5. Use generic media placeholders. Exclude private media, output previews, histories, credentials, |
| cache files, and machine-specific filesystem paths from the public release. |
|
|
| Do not create empty directories for hypothetical future families. Add a family when its first |
| artifact is ready. Published filenames should remain stable; distinguish new versions explicitly |
| when compatibility changes. Existing upstream filenames may retain underscores for loader compatibility. |
|
|
| ## Catalog |
|
|
| `manifests/artifacts.json` uses `schema_version: 1` and an `artifacts` array. |
| Every entry has a stable `id`, `family`, `group`, `kind`, `tasks`, `modalities`, and `files`. |
| Model-specific fields, such as `rank`, `base_model`, or `precision`, are optional for other kinds. |
| Each file record includes its repository-relative path, role, byte size, and SHA-256. |
| Workflow references identify their runtime and format. Compatibility constraints belong to the |
| artifact entry and group documentation, rather than being inferred from a shared repository tag. |
|
|
