LaLa-World/lala-decoder-dinov2-s

LaLa-World pixel decoder (DINOv2 ViT-S/14)

A 12-layer ViT decoder that maps frozen DINOv2 ViT-S/14 patch latents (16 x 16 x 384), including the futures LaLa-World imagines, back to 224 x 224 RGB. Used for visualization only.

file	vision encoder	latents
lala-decoder-dinov2-s.pt	DINOv2 ViT-S/14	16 x 16 x 384

python -m decoder.recon --ckpt lala-decoder-dinov2-s.pt --images a.png b.png --out recon.png
python -m evals.lala_world.run_planning ... --save-pred-decode 4 --decoder-ckpt lala-decoder-dinov2-s.pt

The file holds the EMA and raw decoder weights, the feature RMS used for noise augmentation and the training config. The decoder must match the vision encoder of the LaLa-World / LaLa-Policy checkpoint; use it with lala-world-dinov2-s checkpoints. Decoders for the other encoders: lala-decoder-dinov3-l, lala-decoder-siglip2-so400m.