luuuulinnnn/evs-scene-extrapolation

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.