szlgallen/RnG_model

RnG Model Checkpoints

RAEv2 VGGT MLS K4 Objaverse 512 Scratch16

Path:

pretrained_models/stage1/objaverse-512/vggt-mls-k4-scratch16/

This is the RAEv2 Stage1 decoder pretrained with a frozen official VGGT encoder on the Objaverse multiview data used by RnG-fa3.

Uploaded files:

checkpoints/ep-0000016.pt              Full RAEv2 training checkpoint from epoch 16, step 14832.
stats.pt                              Pre-projection VGGT MLS feature stats, mean/var [2048,32,32].
exports/decoder.pt                    EMA GeneralDecoder-only state_dict, latent dim 768.
exports/projection_plus_decoder.pt    EMA projection + GeneralDecoder state_dict.
config.yaml                           Training config used for this run.
log.txt                               Training log.
README.md                             Detailed artifact-level usage guide.
UPLOAD_MANIFEST.json                  Machine-readable artifact summary.

Recommended RnG-fa3 usage:

exports/projection_plus_decoder.pt + stats.pt

Feature flow:

vggt_features [B,V,1024,2048]
  -> flatten [B*V,1024,2048]
  -> apply stats.pt before projection, reshaped from [2048,32,32] to [1,1024,2048]
  -> Linear(2048 -> 768)
  -> RAE GeneralDecoder
  -> recon [B*V,3,512,512] in [0,1]

Do not apply this stats.pt after projection. It is VGGT pre-projection feature normalization, not DINOv3/RAE latent normalization.

The decoder output is already [0,1] pixel space. Do not apply the old ImageNet denormalization path when using this decoder.