Document precomputed references and processed training data
Browse files
README.md
CHANGED
|
@@ -22,3 +22,17 @@ not standalone Diffusers pipelines or safetensors exports.
|
|
| 22 |
The two JSON manifests identify all eight files with sizes and SHA256 hashes.
|
| 23 |
See [the code and setup guide](https://github.com/JaeyeonKim01/one_step_priming)
|
| 24 |
for model definitions, evaluation, training recipes, and reproduction records.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 22 |
The two JSON manifests identify all eight files with sizes and SHA256 hashes.
|
| 23 |
See [the code and setup guide](https://github.com/JaeyeonKim01/one_step_priming)
|
| 24 |
for model definitions, evaluation, training recipes, and reproduction records.
|
| 25 |
+
|
| 26 |
+
## Evaluation reference assets
|
| 27 |
+
|
| 28 |
+
`assets/fid_stats/` contains the frozen FDr5 reference statistics and the extra
|
| 29 |
+
ConvNeXt statistic used during training diagnostics. The precomputed Inception
|
| 30 |
+
precision/recall reference features are under `data/prc/`. The code's
|
| 31 |
+
`scripts/download_assets.py` downloads the required subset, pins upstream
|
| 32 |
+
evaluation encoders, and fetches the separate official JiT FID reference.
|
| 33 |
+
Users evaluating Priming do not need to recompute real-image statistics.
|
| 34 |
+
|
| 35 |
+
The processed ImageNet training release is in the separate
|
| 36 |
+
[dataset repository](https://huggingface.co/datasets/jaykim1102/priming), with
|
| 37 |
+
ImageNet access terms and matching class indices/OT calibration. Neither
|
| 38 |
+
sampling nor final FID/FDr5 evaluation requires that training dataset.
|