Attribution
Licensing (code and weights) is stated once, in the License section of the model card (README.md).
- DINOv3 pretrained encoder: Meta, official repository, paper, timm model. The transformer blocks are used frozen; the input patch embedding is adapted for this autoencoder. The weights are therefore distributed under the included DINOv3 License; see the model card's License section.
- Encoder implementation: provided by the separately installed
timm dependency, pinned
to
timm==1.0.26. - semantic_vae (well9472/semantic_vae): the design follows its key ideas. These are a pretrained DINO encoder adapted at its edges and a latent split into a reconstruction part and a DINO-derived part. Its encoder has two DINOv2-B branches: one with a trainable patch embedding, whose last six blocks are each LayerNormed and projected to 32 channels, and one fully frozen, whose final block is projected to 32 channels. We also follow it in training the decoder mainly with a DINO feature loss and in using VISReg as the latent regularizer.¹ No code or weights from it are included.
- Decoder up-path: the VQGAN f16 decoder layout of Esser, Rombach and Ommer, Taming Transformers for High-Resolution Image Synthesis, slightly modernized (per-position RMSNorm, folded upsampling) and behind a ViT trunk.
- Earlier decoder head (not in the release): a pixel-level head after PixelDiT (NVlabs/PixelDiT), with the FCDM block of Kwon et al., Reviving ConvNeXt for Efficient Convolutional Diffusion Models, as its token mixer. The released encoder and decoder trunk were trained with it before the convolutional head replaced it.
- Latent regularizer: VISReg (project page), used during training only.
- Optimizer (training only, not a runtime dependency): dionw, a single-GPU implementation of Dion / Dion3 with NorMuon normalization, using kernels from microsoft/dion and Dao-AILab/gram-newton-schulz.
- Training data: about 14 million images from a mix of public and licensed or curated datasets: mostly photographs, plus book covers, a few text-heavy datasets and 1% synthetic rendered text. No training images are redistributed.
¹ semantic_vae's public model card names SIGReg (from LeJEPA) at 1e-3 as its regularizer. We verified in its released training code that the regularizer is VISReg, at 1e-3.