Drifting Models Reproduction (PyTorch)

Community reproduction of Generative Modeling via Drifting in PyTorch.

Project status and claim boundaries

• This repository is not an official release from the paper authors.
• We are actively hardening paper-faithful semantics and evidence artifacts.
• We do not currently claim full paper-level metric reproduction.
• Pixel pipeline remains experimental and should not be treated as parity-closed.

See:

• docs/faithfulness_status.md
• docs/reproduction_report.md
• docs/experiment_log.md
• docs/eval_contract.md

Paper artifact footprint

• The repository tracks the markdown scan and figure assets under Drift_Models/.
• Large convenience artifacts are intentionally untracked and gitignored:
  • Drift_Models.pdf
  • Drift_Models/Drift_Models.json
• Keep those files locally if needed for planner/critique workflows, but do not commit them.

Quickstart (60 seconds)

Option A: uv (recommended)

uv sync --extra dev --extra eval --extra sdvae
uv run python scripts/runtime_preflight.py --device auto --check-torchvision --strict
uv run python scripts/train_toy.py --config configs/toy/quick.yaml --output-dir outputs/toy_quick --device cpu

Option B: pip

python -m venv .venv
source .venv/bin/activate
pip install -U pip
pip install -e ".[dev,eval,sdvae]"
python scripts/runtime_preflight.py --device auto --check-torchvision --strict
python scripts/train_toy.py --config configs/toy/quick.yaml --output-dir outputs/toy_quick --device cpu

Installation guides

• Linux + NVIDIA CUDA: docs/install_linux_cuda.md
• CPU-only: docs/install_cpu_only.md
• macOS (Apple Silicon / MPS): docs/install_macos.md
• Windows + WSL2: docs/install_windows_wsl2.md

Common workflows

• Toy trajectory training: docs/getting_started.md
• Latent smoke training: docs/getting_started.md
• Sampling/eval smoke: docs/getting_started.md
• Full command catalog: docs/commands.md

Compatibility tiers

Compatibility and support policy is documented in:

• docs/compatibility_matrix.md

Runtime health

• Runtime preflight is enforced in CI on Linux/macOS/Windows and nightly on Linux.
• Preflight JSON reports are uploaded as workflow artifacts for each run.
• CI also generates an aggregated runtime summary + failure triage and posts it as a sticky PR comment.
• Runtime diagnostics guide: docs/runtime_health.md
• Local preflight entrypoint: scripts/runtime_preflight.py

Reproducibility and evidence

• Run metadata contracts: docs/provenance_contract.md
• Claim/evidence mapping: docs/claim_to_evidence_matrix.md
• Release parity gate: docs/release_gate_checklist.md
• Public release gate: docs/RELEASE_CHECKLIST.md
• Branch protection policy: docs/branch_protection.md
• PyPI/TestPyPI publish setup: docs/pypi_trusted_publishing.md

Contributing and governance

• Contribution guide: CONTRIBUTING.md
• Code of conduct: CODE_OF_CONDUCT.md
• Security policy: SECURITY.md
• Changelog: CHANGELOG.md

Citation

If you use this repository, cite the original paper and this implementation repo.

Paper: Generative Modeling via Drifting (arXiv:2602.04770).