PDFT-AR adds FD-Loss post-training and evaluation code to four existing image generator codebases. The upstream projects retain their original copyright and license terms.
| Directory | Upstream project | Upstream license |
|---|---|---|
LlamaGen/ |
FoundationVision/LlamaGen | MIT |
VAR/ |
FoundationVision/VAR | MIT |
GigaTok/ |
SilentView/GigaTok | MIT |
TiTok/ |
bytedance/1d-tokenizer | Apache-2.0 |
The FD representation-space loss and reference-statistics workflow follows Jiawei-Yang/FD-Loss. Model checkpoints, reference statistics, and pretrained representation networks are downloaded separately and remain subject to their respective licenses and usage terms.
New PDFT-AR-specific files and modifications are distributed under Apache-2.0.
See LICENSE and LICENSES/Apache-2.0.txt. MIT-derived portions may also be
used under the MIT text in LICENSES/MIT.txt.