YAML Metadata Warning:empty or missing yaml metadata in repo card

Check out the documentation for more information.

SSVTP50x balanced failure run (Round 37/38 negative result)

100k optim steps, batch 8 grad_accum 4, pixel-space DDPM + UNet2DConditionModel + frozen CLIP. Manifest: SSVTP train upsampled 50x + HCT (video-level split).

Result: pixel-space FAILED to add tactile information

Metric 213k baseline SSVTP50x best SSVTP50x final
LPIPS gen vs real 0.377 0.303 0.304
pixel L2 0.136 0.122 0.121
DINO std ratio 1.024 1.165 1.184
V_only acc 65.1% 65.1% 65.1%
V+real_T 69.3% (+4.2pp) 69.3% 69.3%
V+gen_T 65.6% (+0.4pp) 54.0% (-11.1pp) 52.0% (-13.1pp)

Reconstruction quality improved (gen tactile looks more like real) but information content COLLAPSED — gen_T now actively hurts probe accuracy.

Verdict

Pixel-space DDPM trains visual class-prior shortcut. Round 38 pivots to latent vis2tac (Sparsh DINOv2 feature space).

Downloads last month

-

Downloads are not tracked for this model. How to track
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support