README.md

---
library_name: diffusers
license: apache-2.0
datasets:
- laion/relaion400m
base_model:
- black-forest-labs/FLUX.2-dev
tags:
- tae
- taef2
---

# About

Tiny AutoEncoder trained on the latent space of [black-forest-labs/FLUX.2-dev](https://huggingface.co/black-forest-labs/FLUX.2-dev)'s autoencoder. Works to convert between latent and image space up to 20x faster and in 28x fewer parameters at the expense of a small amount of quality. 

Code for this model is available [here](https://huggingface.co/fal/FLUX.2-Tiny-AutoEncoder-FlashPack/blob/main/flux2_tiny_autoencoder.py). Requires [flashpack](https://github.com/fal-ai/flashpack).

# Round-Trip Comparisons

| Source | Image |
| ------ | ----- | 
| https://www.pexels.com/photo/mirror-lying-on-open-book-11495792/ | ![compare_autoencoders_1](https://cdn-uploads.huggingface.co/production/uploads/64429aaf7feb866811b12f73/u7ZnjY8FAwu09-iyEC_um.png) |
| https://www.pexels.com/photo/brown-hummingbird-selective-focus-photography-1133957/ | ![compare_autoencoders_2](https://cdn-uploads.huggingface.co/production/uploads/64429aaf7feb866811b12f73/ZzvJu3VfrzlvZ7bDDASog.png) |
| https://www.pexels.com/photo/person-with-body-painting-1209843/ | ![compare_autoencoders_3](https://cdn-uploads.huggingface.co/production/uploads/64429aaf7feb866811b12f73/B56LPhLYiGT0ffnBVIRbP.png) |

# Usage

```py
import torch
import torchvision.transforms.functional as F

from PIL import Image
from flux2_tiny_autoencoder import Flux2TinyAutoEncoder

device = torch.device("cuda")
tiny_vae = Flux2TinyAutoEncoder.from_pretrained_flashpack(
    "fal/FLUX.2-Tiny-AutoEncoder-FlashPack",
    device=device,
)

pil_image = Image.open("/path/to/image.png")
image_tensor = F.to_tensor(pil_image)
image_tensor = image_tensor.unsqueeze(0) * 2.0 - 1.0
image_tensor = image_tensor.to(device, dtype=tiny_vae.dtype)

with torch.inference_mode():
    latents = tiny_vae.encode(image_tensor, return_dict=False)
    recon = tiny_vae.decode(latents, return_dict=False)
    recon = recon.squeeze(0).clamp(-1, 1) / 2.0 + 0.5
    recon = recon.float().detach().cpu()

recon_image = F.to_pil_image(recon)
recon_image.save("reconstituted.png")
```