# TBSM pretrained assets

This directory contains the pretrained generator backbones and representation
models referenced by the TBSM training configurations. Stable Diffusion VAE
assets are intentionally excluded.

## Naming

Generator checkpoints follow
`{source}_{architecture}_{dataset}_{resolution}_{space}_{role}.pt`.
Representation checkpoints retain the public model family, training dataset,
and input resolution in their names.

## Source mapping

| Release name | Original asset |
| --- | --- |
| `jit_b16_in1k_256_pixel_backbone.pt` | JiT-B/16 ImageNet-1K 256px backbone |
| `jit_b16_in1k_256_pixel_fd_backbone.pt` | JiT-B/16 FD-loss tuning initialization |
| `pixeldit_xl_in1k_256_pixel_ema_backbone.pt` | PixelDiT-XL ImageNet-1K 256px EMA backbone |
| `pixeldit_xl_in1k_512_pixel_ema_backbone.pt` | PixelDiT-XL ImageNet-1K 512px EMA backbone |
| `ucgm_dit_xl2_in1k_256_sdvae_backbone.pt` | UCGM DiT-XL/2 ImageNet-1K 256px SD-VAE backbone |
| `ucgm_dit_xl4_in1k_512_sdvae_backbone.pt` | UCGM DiT-XL/4 ImageNet-1K 512px SD-VAE backbone |
| `mae_vit_b16_in1k_224.pth` | MAE ViT-B/16 ImageNet-1K checkpoint |
| `torchvision_resnet18_in1k_224.pth` | TorchVision ResNet-18 ImageNet-1K checkpoint |
| `siglip2_vit_b16_webli_256/` | SigLIP2 ViT-B/16 WebLI 256px checkpoint and config |

Use `sha256sum -c SHA256SUMS` from this directory to verify the files.