Louym/WM-Prefill-raw-videos

WM-Prefill — raw experiment videos

(中文版: README_zh.md)

The idea in one paragraph. Interactive video world models face a speed/quality dilemma: a bidirectional video model renders a camera trajectory accurately but is far too slow to interact with, while a causal (streaming) model is fast but drifts — as it generates frame after frame it slowly forgets the scene and invents a new one. This project ports the LLM prefill/decode trick to world models: let the slow bidirectional model generate a short exploration once, write its clean latents into the fast causal model's KV cache (its attention memory), then let the fast model generate interactively while attending to that memory. The seeded memory reduces drift by up to +10.9 dB PSNR in the best case, in proportion to how much the memory covers the viewpoints being generated.

This repo holds the raw, full-quality mp4 outputs (832×464, 16 fps) of every experiment. Code, analysis, and the full write-up live in the (public) repo https://github.com/Louym/WM-Prefill — see docs/RESUME.md §7 there for the complete results and history. Models: LingBot-World (Wan2.2 A14B) bidirectional base + its distilled causal fast model.

---

How to watch any demo

One folder = one demo, and every demo folder contains the same four videos:

file	what it is
prefill.mp4	the exploration the fast model was given as memory (generated by the bidirectional model)
bidirectional.mp4	the slow bidirectional model running the demo's decode action — the quality ceiling and the PSNR reference
pure_causal.mp4	the fast causal model running the decode action alone — watch it drift (scenes typically decay into generic terrain, dark blobs, or gravel textures)
ours.mp4	the fast model running the same decode action with its KV cache pre-seeded by prefill.mp4 — the drift is visibly reduced
ours_full.mp4	(only in same_action_replay/benchmark_5s/, same_action_replay/matrix_15s/ridge_turnR/ and different_actions/matrix_15s/ridge_turnR_tour/) the pipeline's complete output: the injected prefill segment, then a hard jump cut back to the input image, then the causal decode. Only the post-cut part is scored; the cut is by design

Always compare within one folder — absolute quality differs between scenes; only within-folder differences are meaningful. In same-variant demos prefill.mp4 is byte-identical to bidirectional.mp4 (the memory is the answer — that's the point of the replay upper bound); it is duplicated so every folder has the same four files. In tour variants the prefill is the scene's 24 s tour.

Prefill variants (what memory the fast model was given):

variant	prefill action	meaning
same	identical to the decode action	replay: the memory literally contains the answer — the upper bound
walk	w-160 (10 s walk forward)	generic, action-agnostic memory
outback	w-80,s-80 (10 s out-and-back)	generic; returns to the start viewpoint
scan	l-40,j-80,l-40 (10 s yaw sweep in place)	covers headings without moving — a deliberate probe
tour / tour24	the scene's ~24 s multi-segment tour	broad spatial coverage without knowing the decode action
mirror30	the 30 s tour with segments reordered	same content, different order (see caveats — this rollout collapsed)

Camera-action notation (used in folder names and above): w/a/s/d = move forward/left/back/right, i/k = pitch up/down, j/l = yaw left/right; letters combine (wl = walk forward while turning); the number = duration in frames at 16 fps (wl-120 = 7.5 s of walking-and-turning). Multi-segment tours are comma-joined.

Measurement. All PSNR numbers below are computed only over the causal decode region (never over the injected prefill frames) against bidirectional.mp4, on raw frames before video encoding. gain = ours − pure_causal.

---

Directory guide & results

The repo has two top-level parts, matching the two scientific questions:

same_action_replay/   the prefill IS the decode action — "does seeded memory help at all?"
                      This is the upper bound: the memory literally contains the answer.
different_actions/    the prefill is a DIFFERENT exploration — "does memory transfer?"
                      The harder, more practical question (you rarely know the future action).

Inside each part, demos are grouped by experiment family (clip length / setting).

Part 1 — same_action_replay/

benchmark_5s/<scene>/ — 8 scenes, ~5 s replays (single GPU):

scene	action	pure	ours	gain
ridge	wj-80	15.66	22.03	+6.37
room	w-80	16.14	17.71	+1.57
industrial	s-80	22.24	23.75	+1.51
reef	a-80	12.69	14.03	+1.34
road	w-80	12.55	13.70	+1.15
street	s-80	21.82	22.64	+0.82
aisle	s-80	13.91	14.43	+0.52
reefcorridor	s-80	14.15	13.75	−0.40

Mean +1.61 dB, 7/8 positive; gains scale with how drift-prone the motion is. Best single watch: ridge/ — pure causal collapses a mountain vista into a generic slope.

tours_24s/<scene>/ — ~24 s multi-segment tours (full attention, multi-GPU):

scene	pure	ours	gain
ridge	13.84	18.75	+4.91
corridor	11.98	14.53	+2.55
room	11.59	13.69	+2.10
road	9.14	10.06	+0.92
reef	12.44	12.35	−0.09
industrial	23.08	22.79	−0.29

The gain scales with drift: big-turn tours that revisit viewpoints gain most; scenes where the causal model barely drifts (pure ≈ 23 dB) have nothing to fix.

matrix_15s/<scene>_<decode>/ — 15 s single-action decodes (fwd / turnL / turnR / back):

demo	pure	gain	demo	pure	gain
ridge_fwd	22.62	+3.00	corridor_fwd	13.33	+4.31
ridge_turnL	14.06	+6.96	corridor_turnL	11.92	+10.94 ⚠
ridge_turnR	12.69	+5.25	corridor_turnR	11.92	+4.75
ridge_back	20.04	+1.51	corridor_back	12.11	+5.67

Continuous turns drift hardest, so replay fixes the most. (⚠ see caveat 2 on the corridor_turnL reference.)

thirty30_30s/ridge/ — the longest run: 30 s prefill + 30 s decode (wl-120,wj-120,i-60,k-60,wl-120; combined KV 241 latent ≈ 298 GB, sequence-parallel): pure 13.92 → ours 16.53, +2.61. The gain persists at 2× the fast model's trained context, but attenuates.

length_sweep_5_15_30s/<scene>_<L05|L15|L30>/ — 4 new scenes × 3 decode lengths, where the three decodes are nested truncations of one trajectory (wl-80 ⊂ wl-160,wj-80 ⊂ wl-160,wj-160,wl-160) — a controlled gain-vs-length test:

scene	5 s	15 s	30 s
city3 (skyline vista)	+3.85	+2.59	+3.36
square44 (monument square)	+2.53	+2.22	+2.27
alpine66 (alpine lake)	+3.10	+1.47	+0.68
cafe15 (café interior)	+0.34	+1.02	+0.14

Length attenuation is scene-dependent, not universal: structured vista/landmark scenes hold their gain at 30 s; the water scene decays; the interior never benefits (ours drifts dark — a valid null). Must-watch: square44_L30/pure_causal.mp4, where the equestrian statue gradually comes alive as a real horse the camera follows — while ours.mp4 holds the square.

Part 2 — different_actions/

The prefill no longer contains the answer. Variants: the 10 s generics (walk, outback, scan), the scene's 24 s tour (spatial coverage without knowing the action), and at 30 s also tour24 and mirror30.

tours_24s/<scene>_<variant>/ — 10 s generic prefills vs the ~24 s tours (gain vs the same pure causal as Part 1):

scene	walk	outback	scan
ridge	+1.17	+1.46	+0.82
corridor	+1.98	+2.10	+0.12
room	+1.29	+0.41	−2.42
road	+0.27	−0.15	+0.31
reef	−1.01	−0.76	−1.21
industrial	−0.57	−0.13	−0.72

A 10-second generic walk recovers ~70 % of the replay gain on the drift-prone scenes; where there is no drift to fix, irrelevant memory mildly hurts; the stationary scan shows viewpoint overlap, not heading coverage, is what transfers.

matrix_15s/<scene>_<decode>_<variant>/ — which prefill for which decode:

ridge	walk	outback	scan	tour
fwd	−0.93	−1.69	−3.16	−2.61
turnL	+1.84	+2.07	+2.44	+2.63
turnR	+2.60	+2.45	+1.31	+2.63
back	−0.95	−0.41	−1.99	−1.86

corridor	walk	outback	scan	tour
fwd	+2.87	+1.07	−1.45	+1.17
turnL	+1.42	+1.63	+0.68	+1.88
turnR	+1.23	+1.34	−0.05	+1.71
back	+1.83	+2.75	+0.09	+2.26

Replay (Part 1) dominates every one of these; the space-covering tour is neither necessary nor sufficient (≈ the 10 s walk on average, at 2.3× the length); prefill only pays where the decode drifts. ridge_turnR_tour/ includes an end-to-end ours_full.mp4 (as does Part 1's ridge_turnR).

thirty30_30s/ridge_<variant>/ — different actions at the 30 s scale:

variant	gain	visual verdict
tour24	+3.88	healthy — beats the 30 s replay (+2.61)
outback	+1.48	holds ~15 s, then collapses like pure causal
walk	+1.31	same pattern
scan	+0.96	same pattern
mirror30	(+1.71)	DEGENERATE — see caveats

Two phenomena visible in the clips: (1) the first case of a different-action prefill beating replay — a shorter, denser memory (214 vs 241 combined latent frames) beats a longer one; (2) the weak arms collapse right after the tour's pitch-up segment — featureless sky strips the causal model of anchors; only memories containing the post-sky vista bridge the gap.

⚠️ Caveats — read before quoting numbers

1. different_actions/thirty30_30s/ridge_mirror30/ours.mp4 is a degenerate rollout: the decode collapses into blue ripple patterns within seconds (the prefill video itself was healthy). Its "+1.71" is a broken-output-vs-drifted-output comparison. Kept for the record; do not report it as a gain.
2. the corridor_turnL demos' bidirectional.mp4 (both parts) is nearly static — the bidirectional model under-executes that turn (inter-frame motion 1.6 vs ~12 for the turnR references). The +10.94 dB cell partly rewards copying a degenerate reference. Within-folder comparisons remain valid; don't showcase that cell. (Turn-right references execute well; turn-left ones poorly — check the reference before using any demo.) The ridge_fwd and ridge_back matrix references are also low-motion.
3. Some source images carry a small on-screen keyboard HUD (part of the benchmark input images); it is faithfully reproduced by the reference and is not an artifact.
4. Every experiment here used seed 42; treat single-cell numbers as observations, means and repeated patterns as findings.

Reproduction

Generation scripts, measurement code, the attention-probe analysis, and interactive findings pages are in https://github.com/Louym/WM-Prefill (start at docs/RESUME.md). PSNR was computed on pre-encode frames; these mp4s are near-lossless encodes of the same frames. This dataset is private; the same content is mirrored on GitHub (Louym/WM-Prefill-raw-videos) and Hugging Face (datasets/Louym/WM-Prefill-raw-videos).