gatilin/open-genception-14b-pytorch

Open-GenCeption 14B PyTorch (Diffusers)

Format conversion of Google DeepMind's released GenCeption Colab checkpoint from JAX/NNX .npz to a Diffusers WanTransformer3DModel safetensors file.

This is the multi-task author implementation, not Specialist-S and not a Table 1 / Table 2 checkpoint. paper_comparable=false.

Upstream	https://github.com/google-deepmind/representations4d (Apache-2.0)
Source weights	https://storage.googleapis.com/representations4d/checkpoints/genception/genception_14b_transformer.npz
Conversion	https://github.com/isLinXu/open-genception scripts/export_author_impl_pytorch.py
DiT	14.288B params, 1095 tensors, bfloat16, 40 layers, 40 heads, FFN 13824
Token heads	12 tensors in token_heads/ (token1 13×1, token2 33×3)
VAE	Official PyTorch WAN VAE copied from the DeepMind release. Not a JAX conversion.

The DiT loads with Diffusers. Token heads are a sidecar: they need the grouped attention path in Open-GenCeption and are not a drop-in Wan pipeline.

import torch
from diffusers import AutoencoderKLWan, WanTransformer3DModel

dit = WanTransformer3DModel.from_pretrained(
    "gatilin/open-genception-14b-pytorch",
    subfolder="transformer",
    torch_dtype=torch.bfloat16,
)
vae = AutoencoderKLWan.from_pretrained(
    "gatilin/open-genception-14b-pytorch",
    subfolder="vae",
    torch_dtype=torch.bfloat16,
)

Provenance, key counts, and the source .npz sha256 are in export_ledger.json.