About

Tiny AutoEncoder trained on the latent space of 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.

Round-Trip Comparisons

Example Usage

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(
    "fal/FLUX.2-Tiny-AutoEncoder",
).to(device=device, dtype=torch.bfloat16)

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")

Use with Diffusers 🧨

import torch
from diffusers import AutoModel, Flux2Pipeline

device = torch.device("cuda")
tiny_vae = AutoModel.from_pretrained(
    "fal/FLUX.2-Tiny-AutoEncoder", trust_remote_code=True, torch_dtype=torch.bfloat16
).to(device)

pipe = Flux2Pipeline.from_pretrained(
    "black-forest-labs/FLUX.2-dev", vae=tiny_vae, torch_dtype=torch.bfloat16
).to(device)