---
base_model: stabilityai/stable-diffusion-3.5-medium
library_name: diffusers
pipeline_tag: text-to-image
tags:
- lora
- diffusionnft
- baseline
- training-logs
---

# DiffusionNFT baselines

Logs and LoRA checkpoint artifacts for **baseline experiments only**.
Rank-aware algorithm experiments are excluded.

Included runs:

- `nft_sd3_geneval_2026.04.20_11.18.57`
- `nft_sd3_geneval_2026.04.22_04.13.06`

The archive contains 9 checkpoint directories and
703 source files (2.489 GiB before
media compression; 2.488 GiB of published payload).
These are adapters for Stable Diffusion 3.5 Medium, not the base model weights.

## Provenance and checkpoint ownership

`experiment_provenance.csv` records the selected runs' saved Git commits,
configurations, and baseline classification.

The April 20 and April 22 runs record the upstream NVlabs/DiffusionNFT commit
`cbb14f84b8312b620390dfcbe2fab69c1383b104`, with principal hyperparameters matching
the upstream configuration. No working-tree snapshot or uncommitted diff was
saved, so a byte-for-byte pristine source tree cannot be proved.

Both runs recorded `logs/nft/sd3/geneval` as their checkpoint save directory.
That directory was later reused by the April 23 run. The timestamps of
`logs/nft/sd3_v0/geneval/checkpoints/checkpoint-0` through `checkpoint-240`
(every 30 steps) align with the April 20 run, indicating a likely archived copy.
There is no explicit move record. These artifacts retain their original paths.
**No separately identifiable April 22 checkpoints were found.**

## Contents

- `logs/nft_sd3_*/wandb/run-*/`: W&B binary records, configs, summaries, output
  logs, environment records, and generated-media archives for the selected runs.
- `logs/nft/sd3_v0/geneval/`: the April 20-associated checkpoint archive,
  including LoRA, old-policy LoRA, optimizer state, and scaler state.
- `artifact_manifest.jsonl`: payload paths, sizes, SHA-256 values, and source-file counts.
- `SHA256SUMS`: checksums for all payload files.
- `symlink_aliases.json`: original convenience aliases; target files are included
  once instead of duplicating `latest-run` and debug-log aliases.

## Download and extract

Use the Hugging Face CLI or `huggingface_hub` to download selected files.
Generated images are preserved losslessly in `media.tar.zst` next to their
original parent directory. With GNU tar and zstd installed:

```bash
sha256sum -c SHA256SUMS
tar --zstd -xf path/to/media.tar.zst -C path/to/parent
```

Run checksum verification from the repository root. Original local files were
not modified. Credential-pattern scans were performed before publication;
redaction counts, if any, are recorded in the manifest.

Upstream code: https://github.com/NVlabs/DiffusionNFT

Base model: https://huggingface.co/stabilityai/stable-diffusion-3.5-medium