OzzyGT/taew2_1

TAEW 2.1 (mirror)

An unmodified mirror of taew2_1.safetensors from madebyollin/taehv, the tiny autoencoder for the Wan 2.1 latent space. It decodes a latent to an image about fifty times cheaper than the full VAE, which is what makes per-step previews affordable.

Krea 2 and Qwen-Image share Wan 2.1's latent space exactly, so this decodes their latents too. That is why this mirror exists: OzzyGT/krea2_preview_blocks fetches it at runtime to preview Krea 2's prediction while it denoises.

Use

Needs taehv.py from the upstream repo:

from huggingface_hub import hf_hub_download
from taehv import TAEHV

tae = TAEHV(hf_hub_download("OzzyGT/taew2_1", "taew2_1.safetensors"), arch_name="taew2_1").cuda().eval()
frames = tae.decode_video(latents)  # (N, T, C, H, W) normalized latents -> (N, T, 3, H*8, W*8) in [0, 1]

Pass arch_name explicitly: upstream guesses the architecture from the checkpoint's filename, which is a cache path here.

Provenance

• Source: safetensors/taew2_1.safetensors at commit 0ad83bb
• sha256 04766eac0221b5390b985ae3fdcca652cbb4b1e8b82b28ea7ff89dfad1b1a93f
• 22,642,902 bytes, 128 tensors, fp16, 11.3M parameters
• Byte-for-byte the upstream file. Nothing was converted, quantized or retrained.

License

MIT, (c) 2025 Ollin Boer Bohan. The upstream LICENSE is included in this repo unchanged. All credit for the model goes to the author; this repo only hosts a copy on the Hub, since upstream distributes these weights through GitHub.