RAEv2 DINOv3-L K7

Self-contained RAEv2 stage-1 model with the pretrained DINOv3-L encoder and general-domain DINOv3-L K7 decoder. Latent shift/scale statistics are intentionally not included.

import torch
from transformers import AutoModel

model = AutoModel.from_pretrained(
    "YOUR_ORG/YOUR_REPO",
    trust_remote_code=True,
).eval().to("cuda")

# RGB float tensor in [0, 1] (a [0, 255] tensor is also accepted).
images = torch.rand(1, 3, 256, 256, device="cuda")
latents = model.encode(images)       # [1, 1024, 16, 16]
reconstructions = model.decode(latents)  # [1, 3, 256, 256]

# The components are directly accessible too.
latents = model.encoder(images)
reconstructions = model.decoder(latents)

decode returns the decoder's raw RGB prediction; it does not clamp or rescale the result.

Requires transformers>=4.56.