You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
PRIMA creates a 3D quadruped mesh from a single 2D image. It leverages BioCLIP-based biological priors for robust cross-species shape understanding, then applies test-time adaptation with 2D reprojection and auxiliary keypoint guidance to refine SMAL pose and shape predictions.
17
17
18
18
It further can be used to build Quadruped3D, a large-scale pseudo-3D dataset with diverse species and poses.
@@ -21,15 +21,19 @@ PRIMA achieves state-of-the-art results on Animal3D, CtrlAni3D, Quadruped2D, and
21
21
22
22
## Installation
23
23
24
+
PRIMA requires Python 3.10 or newer. A CUDA-enabled PyTorch installation is
25
+
recommended for local inference and training.
26
+
24
27
### Install from PyPI
25
28
26
-
> Recommended: Python 3.10 and a CUDA-enabled PyTorch installation.
29
+
Create a clean environment, install PyTorch for your CUDA version, then install
30
+
the package:
27
31
28
32
```bash
29
33
conda create -n prima python=3.10 -y
30
34
conda activate prima
31
35
32
-
#Install PyTorch matching your CUDA (example: CUDA 11.8)
36
+
#Example for CUDA 11.8. Adjust this command for your CUDA version.
Use these when developing from a **git clone** (not the PyPI wheel). The shell scripts are **non-interactive** (pip uses `--no-input`; `GIT_TERMINAL_PROMPT=0` for git). Put Hugging Face credentials in your environment or git credential helper before pushing the Space.
64
+
The helper script below creates a fresh virtual environment, installs runtime
65
+
dependencies, pulls Git LFS assets if available, downloads the default demo
66
+
checkpoints/data, and verifies that the demo dependencies can be imported:
-`PRIMA_VENV=.venv ./scripts/clean_install_local.sh --skip-data` — skip the large `setup_demo_data` download if `data/` is already populated.
64
-
-`./scripts/clean_install_local.sh --wipe-data --force-data` — delete downloaded `data/` assets and redownload.
65
-
-`./scripts/clean_install_local.sh --no-editable` — only `requirements.txt` (no `pip install -e .`); use if editable install fails and you will install the training stack via conda as in the PyPI section above. You still need **Python 3.10+** for Gradio 5.1+. The smoke test sets `PYTHONPATH` to the repo root so `import prima` works without an editable install.
for the SuperAnimal PyTorch API. On macOS, `clean_install_local.sh` installs
68
-
it separately after a compatible PyTables wheel (`tables>=3.9.2,<3.11`) to
69
-
avoid Apple Silicon build issues. Validate the local setup with
70
-
`./scripts/test_local_full.sh`.
78
+
-`--skip-data` skips the large demo data download if `data/` is already populated.
79
+
-`--wipe-data --force-data` removes downloaded demo assets and downloads them again.
80
+
-`--no-editable` installs dependencies without registering the repo as an editable package.
71
81
72
-
After `requirements.txt`, the script runs **`pip install --no-deps -e .`** so the `prima` package is registered without re-resolving `pyproject.toml` (which would pull **Detectron2** from git again). Install Detectron2 separately if needed:`pip install 'git+https://github.com/facebookresearch/detectron2.git'`.
82
+
On macOS, install Python 3.10 if needed:
73
83
74
-
**Hugging Face Space (full redeploy from your working tree):**
0 commit comments