MCSDF ablation — live results

forward ctx18 (observe 18 → generate 63) · VGGT geometry · no memory slots · current: step ~400 / 10k (4%)

Headline (already visible at 4% training): the forest hallucination is GONE in all three variants — the far/turn frames now generate corridor structure. Since all three share the fixed skeleton (VGGT consistent geometry + real-extrapolation ctx18 cache) and differ only in the timestep field, this cleanly attributes content correctness to memory quality, not the noise mechanism. Ranking A vs B vs baseline is too early — first formal comparison at the 3k matched snapshot.

🎬 slow-mo 5-panel (6 clips)🔬 all-signals 3×3 videos📐 design diagram

Before vs now — far frame (40), pier clip

**PRE-FIX** (shattered mono-depth cache): hallucinated forest · A / B / **baseline** @step400 on the fixed cache: corridor structure, still noisy (undertrained, 4%).

The 4 conditions

condition	training field	inference field	scale
Warp baseline	uniform (mean-matched, no noise map)	uniform SDEdit	LoRA-128 / 10k
A-cont	AsyncPatch random-patch (continuous, task-agnostic)	continuous τ=1−λk	LoRA-64 / 10k
A-mask	(same Model A checkpoint)	binary coverage (covered→0.45 held, unseen→1→0)	LoRA-64 / 10k
B	geometry noise-map (τ ceiling from point-cloud confidence)	continuous τ=1−λk	LoRA-64 / 10k

A-cont vs B = training recipe · A-mask vs A-cont = inference granularity (same weights) · A-mask vs B = full paradigm · baseline (bigger rank!) vs B = does the noise map beat plain warp even at a capacity disadvantage.

Early strips @step400 — frames 0/20/40/60/80

pier / covered walkway — A / B / baseline (top→bottom)

Design

3-run decomposition: 2 trainings → 3 inference conditions; everything else identical.

Milestones: 3k matched snapshot → compare_3way videos (GT | A-cont | A-mask | B) · 10k final. This page refreshes in place.