Instructions to use madebyollin/taef1 with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- Diffusers
How to use madebyollin/taef1 with Diffusers:
pip install -U diffusers transformers accelerate
import torch from diffusers import DiffusionPipeline # switch to "mps" for apple devices pipe = DiffusionPipeline.from_pretrained("madebyollin/taef1", dtype=torch.bfloat16, device_map="cuda") prompt = "Astronaut in a jungle, cold color palette, muted colors, detailed, 8k" image = pipe(prompt).images[0] - Notebooks
- Google Colab
- Kaggle
TAEF1 1.1
Could you provide some info or compare with 1.0?
TAEF1 1.1 "Festive Maple Latte" edition improves the encoder and decoder quality over TAEF1 1.0.
The TAEF1 1.1 encoder was trained longer with a new feature-matching loss, substantially improving TAEF1 compatibility with the FLUX.1 decoder.

The TAEF1 1.1 decoder was also trained longer and with a larger discriminator, slightly improving the quality of TAEF1-decoded images.

The TAEF1 1.1 decoder and encoder improvements result in better roundtrip reconstruction quality. COCO-Val rFID using TAEF1 1.1 for both enc/dec has improved from 11.1 in the initial TAEF1 release (comparable to TAESDXL) to 4.5 (comparable to full SDXL VAE, see https://arxiv.org/abs/2307.01952 table 3).
VAE COCO-Val rFID (run in fp16) TAEF1 1.0 11.1 TAEF1 1.1 4.5 FLUX.1 1.4 That said, TAEF1 1.1 is still lower-quality than the official FLUX.1 VAE and still struggles with challenging, high-detail images.
