dREPA_collections

mkdir -p data/coco256_features_sdvae_ft_ema/train for f in data/coco256_features_sdvae_ft_ema/train*.tar; do tar xf "$f" -C data/coco256_features_sdvae_ft_ema/train/ done # --- Unpack COCO256 val --- mkdir -p data/coco256_features_sdvae_ft_ema/val for f in data/coco256_features_sdvae_ft_ema/val*.tar; do tar xf "$f" -C data/coco256_features_sdvae_ft_ema/val/ done # --- Unpack metrics data --- mkdir -p metrics/data tar xf metrics/data.tar -C metrics/data/ # --- (Optional) Remove tar files after unpacking --- rm -f data/coco256_features_sdvae_ft_ema/train*.tar rm -f data/coco256_features_sdvae_ft_ema/val*.tar rm -f metrics/data.tar ``` Or use the one-liner: ```bash cd dREPA_collections && bash unpack.sh ``` ### 3. Verify After unpacking, the directory should look like: ``` data/coco256_features_sdvae_ft_ema/ ├── empty_context.npy ├── train/ │ ├── 0.png, 0.npy, 0_0.npy, 0_1.npy, ... # 82,783 samples │ └── ... └── val/ ├── 0.png, 0.npy, 0_0.npy, 0_1.npy, ... # 40,504 samples └── ... ``` Each sample consists of: - `{idx}.png` — original image (256x256) - `{idx}.npy` — SD-VAE latent features - `{idx}_{k}.npy` — text embeddings (k=0..4, multiple captions per image) ### 4. Link to dREPA project ```bash # Assuming dREPA repo is at ./dREPA ln -s $(pwd)/dREPA_collections/pretrained_models/* dREPA/pretrained_models/ ln -s $(pwd)/dREPA_collections/dinov3 dREPA/dinov3 ln -s $(pwd)/dREPA_collections/eval_references/VIRTUAL_imagenet256_labeled.npz dREPA/ # Set data dir in training scripts export DATA_DIR=$(pwd)/dREPA_collections/data/coco256_features_sdvae_ft_ema # for MMDiT export DATA_DIR=$(pwd)/dREPA_collections/data/imagenet256 # for SiT-256 ``` ## Data Formats | Dataset | Format | Used by | Size (approx) | |---------|--------|---------|---------------| | coco256_features_sdvae_ft_ema | .png + .npy | MMDiT (train_mmdit.py) | ~50-80 GB | | imagenet256 | HF Arrow | SiT-256 (train_sit.py) | ~150 GB | | imagenet512 | HF Arrow | SiT-512 (train_sit.py) | ~620 GB | ## Pretrained Encoder Weights | File | Model | Size | |------|-------|------| | dinov2_vitb14_pretrain.pth | DINOv2 ViT-B/14 | 331 MB | | dinov3_vitb16_pretrain_lvd1689m-*.pth | DINOv3 ViT-B/16 | 327 MB | | dinov3_vitl16_pretrain_lvd1689m-*.pth | DINOv3 ViT-L/16 | 1.1 GB | | dinov3_vith16plus_pretrain_lvd1689m-*.pth | DINOv3 ViT-H/16+ | 3.1 GB | | dinov3_vit7b16_pretrain_lvd1689m-*.pth | DINOv3 ViT-7B/16 | 25 GB | | mocov3_vitb.pth | MoCo v3 ViT-B | 823 MB | | mocov3_vitl.pth | MoCo v3 ViT-L | 2.4 GB | | mae_vitl.pth | MAE ViT-L | 1.1 GB | | ijepa_vith.pth | I-JEPA ViT-H | 9.6 GB | ## License Please respect the original licenses of the pretrained models and datasets: - DINOv2/DINOv3: Meta Platforms (Apache 2.0) - MoCo v3, MAE, I-JEPA: Meta Platforms - ImageNet: Academic use only (ILSVRC license) - MSCOCO: Creative Commons Attribution 4.0