File size: 3,091 Bytes
b4a0ddd | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 | # EVS — Extrapolative Visual Sensing with Geometry-Aware Sensor Tokens
A generative **video diffusion** model for **posed scene extrapolation**: given a short observed video
and a target camera trajectory, generate the video along that trajectory — extrapolating into unobserved
regions while staying **consistent on revisit** (when the camera returns to previously-seen content).
Built on **Wan2.1-VACE-14B**, trained with **Diffusion Forcing**, and conditioned via **Plücker camera
control** + a **geometry-aware "sensor token"** design.
## Method
**Diffusion Forcing (in-context memory).** Independent per-frame noise levels; the observed frames sit in
the same latent sequence at noise ≈ 0 (clean) while target frames are denoised. Memory is *in-context* — the
model's self-attention reads the clean observed frames; no separate memory module.
**Sensor token (the contribution).** Each token carries, beyond its content latent:
1. **Ray position embedding** — the token's world-frame Plücker ray, added to self-attention Q/K, so tokens
that view the same content (e.g. a revisited camera pose) retrieve each other *by geometry*.
2. **Evidence weight** — a depth-free **depth-hypothesis co-visibility** count (how many frames observe the
token's content), injected as an attention key-boost + a content channel (and a knownness prior).
**Camera control.** PermaVid-style Plücker `SimpleAdapter`, additive at the patch embedding (not via the
VACE control branch). Per-frame timesteps require no architecture surgery (Wan `DiTBlock` already accepts
per-token modulation). All new modules are zero-init, so an untrained model equals the base Wan T2V model.
## Data
**RealEstate10K** (static real-estate walkthroughs, smooth cinematic camera). Clips are built as smooth
**out-and-back** trajectories over consecutive frames (forward then retrace) for a genuine revisit with real
GT, plus a **forward-split** eval mode (observe a prefix, extrapolate the forward continuation into unseen).
Geometry (ray codes + evidence) is a pure function of the poses and is precomputed/cached.
## Repo layout
```
code/
evs_model.py sensor-token model (ray-PE + evidence + camera + Diffusion Forcing)
cache_re10k.py RealEstate10K -> cached clips (latents + poses + ray + evidence)
cache_smooth.py DL3DV smooth out-and-back builder (earlier data source)
train_df.py FSDP trainer (per-frame DF noise, target-masked flow-matching, resume + safe save)
infer_df.py Diffusion-Forcing sampling (clean context + generated target)
make_*_viz.py qualitative viz (revisit / data / forward-split comparisons)
make_eval_metrics.py forward-split extrapolation metrics (PSNR/SSIM on generated frames)
*.sbatch, evs_fsdp*.yaml SLURM + accelerate-FSDP launch configs
docs/
DESIGN.md, BUILD_LOG.md design rationale + running build log
```
## Status
Work in progress. Model + pipeline implemented and verified; training on RealEstate10K. Checkpoints and
qualitative samples: see the companion HuggingFace repo.
|