Wan2.2 VAE — Tactile Reconstruction

Source: motherboard/2026-05-10/episode_000/tactile_left.mp4 · Clip: 256×256 · 81 frames · 16fps · Adapter: lora_best.pt (r16 α16 residual_attn, step 4000)

Reconstruction Comparison

originalresized input

reference clip

base Wan2.2frozen VAE

MSE 2.90e-5 · PSNR 45.42 dB

LoRA finetunedtactile adapter

MSE 2.74e-5 · PSNR 45.66 dB · +0.24 dB vs base

diff ×20|original − LoRA|

error heatmap · brighter = larger deviation (amplified 20×)

Cosmos CV4x8x8encode→decode

MSE 4.08e-5 · PSNR 43.89 dB · −1.77 dB vs LoRA

diff ×20|original − Cosmos|

error heatmap · brighter = larger deviation (amplified 20×)

Source: vm_diffusion/runs_mot/my_mot_run_0702/inference_rollout_mot/sample_002/tr_gt.mp4 · Clip: 256×256 · 13 frames · 6fps · Adapter: lora_best.pt (r16 α16 residual_attn, step 4000)

Reconstruction Comparison

originalresized input

reference clip

base Wan2.2frozen VAE

MSE 2.40e-5 · PSNR 46.20 dB

LoRA finetunedtactile adapter

MSE 2.17e-5 · PSNR 46.64 dB · +0.43 dB vs base

diff ×20|original − LoRA|

error heatmap · brighter = larger deviation (amplified 20×)

Cosmos CV4x8x8encode→decode

MSE 2.93e-5 · PSNR 45.33 dB · −1.31 dB vs LoRA

diff ×20|original − Cosmos|

error heatmap · brighter = larger deviation (amplified 20×)