yxma/vtwm-tactile-latent-delta-viz

VideoVAE+ vs Cosmos — Latent tokenizer diagnostic for visuo-tactile world model

Timeline: 2026-07-02 to 2026-07-03. Diagnosing whether our current VideoVAE+ latent tokenizer has architectural issues that hurt downstream world model training, and whether NVIDIA's Cosmos Tokenizer CV4x8x8 is a viable replacement.

TL;DR

Two architectural issues found in VideoVAE+, both absent in Cosmos:

1. Right-bottom spatial hotspot — Downsample2plus1D uses asymmetric zero-pad on right + bottom (autoencoder2plus1d_1dcnn.py:396), accumulated over 3 downsample layers, produces a systematic (15, 15) activation on uniform mid-gray inputs. Content-dependent via GroupNorm suppression → delta-to-ref does NOT fully cancel it.

2. Temporal boundary bleed (ti=0 has 5× baseline magnitude) — spatial encoder's TemporalAttention uses RelativePosition embeddings with full-bidirectional attention (no causal mask) + EncoderTemporal1DCNN front- only replicate pad accumulates over 2 downsample layers, biasing ti=0 by 5× and ti=1 by 1.7× on identical inputs. This is also a training-time information leak: history latents encode future raw frames through bidirectional attention.

Cosmos verification: CausalConv3d uses strict left-only replication pad; CausalTemporalAttnBlock uses lower-triangular mask. For 16 identical raw frames, Cosmos produces 5 latent frames identical to 4 decimal digits (drift mean = 0.0000, ~bf16 precision noise).

Cosmos reconstruction beats VideoVAE+ on tactile by 6–8 dB PSNR even though Cosmos is zero-shot on tactile and VideoVAE+ was finetuned on our exact tactile distribution. Visual: VideoVAE+ leads by ~3 dB.

Recommendation: switch tokenizer to Cosmos CV4x8x8. No tactile finetune needed. Visual gap is acceptable and can be closed later if it becomes a world-model bottleneck.

---

Experiments (chronological)

#	Directory	What	Key finding
01	01_tactile_delta_heatmaps_p01_vs_p01rand/	Tactile Δ heatmaps on 5 episodes × 2 sides × 2 ref modes	Contact regions concentrate in 3-6 latent tokens; p01rand pool covers 8-19% of frames
02	02_videovae_tactile_debug/	VideoVAE+ tactile-finetuned encoder — synthetic input / per-row vmax / temporal identity	gray-192 argmax = (15,15); ti=0 max = 5× baseline; ti=1 max = 1.7×
03	03_videovae_visual_debug_base_vae/	Base VideoVAE+ (visual) — same 3 tests	Same artifacts as tactile — confirmed architectural, not finetune-specific
04	04_videovae_full_latents_tactile_ep000/	Every latent frame of ep_000 as MP4 (tactile)	3-panel per-frame video: raw + |z| global vmax + |z| per-frame vmax
05	05_videovae_full_latents_visual_ep000/	Every latent frame of ep_000 as MP4 (visual_left, base VAE)	Same layout for direct A/B viewing with tactile
06	06_cosmos_debug/	Cosmos CV4x8x8 causality verification	drift mean = 0.0000 on 16 identical frames; hotspot much weaker
07	07_cosmos_vs_videovae_recon/	encode+decode reconstruction A/B on all 5 streams	Cosmos beats VideoVAE+ on tactile by +6-8 dB PSNR (zero-shot); VVAE+ leads visual by +3 dB
08	08_cosmos_full_latents_ep000/	Cosmos full-episode latent videos + p01 Δ heatmaps (all 5 streams)	Cosmos latent 5-7× smaller magnitude, cleaner spatial distribution

Code

All scripts committed to scripts/ in the training repo:

• viz_tactile_latent_delta.py — experiment 01
• debug_tactile_latent_bias.py — experiment 02
• debug_visual_latent_bias.py — experiment 03
• viz_all_latents_video.py — experiments 04 + 05
• debug_cosmos_latent.py — experiment 06
• compare_recon_cosmos_vs_videovae.py — experiment 07
• viz_cosmos_episode.py — experiment 08

Reconstruction metrics summary

From experiment 07, on episode_000 raw frames [1000, 1016):

stream	Cosmos PSNR	Cosmos SSIM	VideoVAE+ PSNR	VideoVAE+ SSIM
visual_left	30.04	0.892	32.99	0.949
visual_middle	29.94	0.885	33.20	0.945
visual_right	28.82	0.889	32.10	0.948
tactile_left	43.98	0.971	37.42	0.965
tactile_right	46.68	0.983	38.83	0.979

Next steps

• Robustness check across more episodes / segments
• Full-dataset re-encoding with Cosmos if the finding holds
• Smoke train on Cosmos latents to check world-model rollout quality