QwenImage VAE distributed with InternVL-U
This package preserves the separately released vae/diffusion_pytorch_model.safetensors component from the official InternVL-U checkpoint. It is an AutoencoderKLQwenImage VAE with a 16-channel latent representation. InternVL-U's VLM and its MMDiT generation decoder are outside this package.
Provenance and license
- Source: InternVL-U/InternVL-U at fixed revision
f012d760e69712bb47f7d3d09a24280f346cee01. - Source files:
vae/config.jsonandvae/diffusion_pytorch_model.safetensors. - The official model repository includes an MIT
LICENSE, retained here.
Load
from huggingface_hub import snapshot_download
import sys, torch
root = snapshot_download("gatilin/internvlu_qwenimage_vae")
sys.path.insert(0, root)
from component import load_model
vae = load_model(root, device="cpu", dtype=torch.float32)
The native interface accepts [batch, 3, frames, height, width]. encode(x).latent_dist exposes the latent distribution and decode(latent).sample returns a reconstruction.
Validation boundary
This release checks exact source SHA-256, strict state-dict loading, and a minimal CPU video encode/decode with finite outputs. These checks verify the component interface only; they do not measure visual quality or evaluate full InternVL-U generation, editing, understanding, or reasoning.