README.md

---
license: mit
language:
- en
tags:
- diffusers
- autoencoder
- vision-foundation-model
- dinov2
- dinov3
- mae
- siglip2
- feature-extraction
- pae
library_name: diffusers
pipeline_tag: feature-extraction
---

# PAE Diffusers Checkpoints

Converted PAE (Prior-Aligned Autoencoder) tokenizer checkpoints in standard Hub custom-pipeline layout.

PAE is a **VAE-free** latent framework. Each variant splits into dedicated components:

| Variant | VFM backbone (decoder config) | Latent dim | Input size |
|---------|------------------|------------|------------|
| `pae-dinov2-large-d32` | DINOv2-L (with registers) | 32 | 224 |
| `pae-dinov3-large-d32` | DINOv3-ViT-L/16 | 32 | 256 |
| `pae-mae-large-d32` | MAE-L | 32 | 256 |
| `pae-siglip2-so400m-d32` | SigLIP2-SO400M | 32 | 256 |

Each variant directory is a self-contained Diffusers repo. Each component subfolder ships **one Python file**:

```text
model_index.json
pipeline.py
scheduler/scheduling_flow_match_pae.py
transformers/transformer_lightning_dit.py
decoder/decoder_pae.py
decoder/diffusion_pytorch_model.safetensors
```

## Usage

```python
from pathlib import Path
import torch
from diffusers import DiffusionPipeline

model_dir = Path("/home/czy/local/models/BiliSakura/PAE-diffusers/pae-dinov2-large-d32").resolve()
pipe = DiffusionPipeline.from_pretrained(
    str(model_dir),
    local_files_only=True,
    custom_pipeline=str(model_dir / "pipeline.py"),
    trust_remote_code=True,
    torch_dtype=torch.bfloat16,
).to("cuda")

print(pipe.get_label_ids("golden retriever"))
```