Wan2.2 tactile VAE · reference-flow LoRA
Reconstruction on held-out motherboard clips
Each GelSight tactile clip is encoded and decoded through the
reference-flow LoRA adapter (lora_out_refflow/lora_last.pt, r32 α32,
trained at 6fps with the L1 loss weighted by reference optical-flow magnitude). Every tab
shows the original , the reconstruction , and their amplified difference .
Reconstruction runs at 6fps — the rate the adapter trained at.
ep002 · tactile_right (10–30s) ep002 · tactile_left (6–26s) ep003 · tactile_left (full 4s)
⏯ Play / pause visible
Source: episode_002/tactile_right.mp4
· 117 frames · 6fps · 10–30s
(stride 5 from 600 native @ 30fps)
original resized + 6fps input
reference · 117 frames @ 256×256
reconstructed reference-flow VAE
PSNR 42.60 dB · MSE 5.50e-05 · base 42.19
difference |original − recon| ×20
hot colormap · brighter = larger error
Source: episode_002/tactile_left.mp4
· 117 frames · 6fps · 6–26s
(stride 5 from 600 native @ 30fps)
original resized + 6fps input
reference · 117 frames @ 256×256
reconstructed reference-flow VAE
PSNR 42.24 dB · MSE 5.97e-05 · base 41.78
difference |original − recon| ×20
hot colormap · brighter = larger error
Source: episode_003/tactile_left.mp4
· 21 frames · 6fps · first 4s
(stride 5 from 120 native @ 30fps)
original resized + 6fps input
reference · 21 frames @ 256×256
reconstructed reference-flow VAE
PSNR 38.27 dB · MSE 1.49e-04 · base 37.55
difference |original − recon| ×20
hot colormap · brighter = larger error
Adapter: LoRA on the Wan2.2 causal 3D VAE (Wan-AI/Wan2.2-TI2V-5B), r32 α32, all
conv targets, 4000 steps @ 6fps, edge+LPIPS+reference-flow-weighted L1. Diff = |original −
recon| through a hot colormap. These clips have no optical-flow sibling, so no motion split.