convert_to_onnx.py

from diffusers import AutoencoderTiny

from pathlib import Path

from optimum.exporters.onnx import export
from optimum.exporters.onnx.model_configs import VaeDecoderOnnxConfig, VaeEncoderOnnxConfig
taesd = AutoencoderTiny.from_pretrained("madebyollin/taesd")

# TAESD Decoder

taesd.forward = lambda latent_sample: taesd.decode(x=latent_sample)
export(model = taesd, config = VaeDecoderOnnxConfig( config = taesd.config, task = "semantic-segmentation"), output = Path("./vae_decoder/model.onnx"))

# TAESD Encoder

taesd.forward = lambda sample: {"latent_sample": taesd.encode(x=sample)["latents"]}
export(model = taesd, config = VaeEncoderOnnxConfig( config = taesd.config, task = "semantic-segmentation"), output = Path("./vae_encoder/model.onnx"))