Component Matrix — every experiment, ✓ / ✗ audited from its sbatch

2026-07-29 · every row verified against the actual training command (not memory) · columns = the pipeline ingredients · ✓ has✗ doesn't✓ had, then removedvalue
experimentbackbonelineagetrainable warp ch
(z_mem+k)
+Plücker
ray
VACE
native
seq-ctx
(attn cond)
mem
slot
k→τ fieldτ mapping z_mem
anchor
per-tok
AdaLN✎
rehearsalliftpatch
fields
sync
slices
rand
(AsyncP)
SYNC
infer
cluster-τ steps / statedev avg
PSNR
— retired / control arms (I2V-14B) —
baseline r128I2V-14BscratchLoRA-128 ✗ uniformflat (mean) 3000 / chain12.88
expAI2V-14BscratchLoRA-64 Ⓑ .45 floor 7500+ / c312.62
expBI2V-14BscratchLoRA-64 Ⓑ .45 floor 7500+ / c312.60
expCI2V-14Bwarm(B@1500)LoRA-64 ✓17✓→del Ⓒ .85 clamp ✓.15 ✗ legacy4500 / frozen12.25
expE-geoI2V-14Bwarm(C@3500)LoRA-64 ✓17 Ⓒ .85 ✓.10✓.35 own~3200 / c212.95
expE-randI2V-14Bwarm(C@3500)LoRA-64 ✓17 Ⓒ .85 ✓.10 own~3200 / c212.93
— main line (I2V-14B, clean lineage) —
expE-scratchI2V-14BscratchLoRA-64 ✓17 Ⓒ .85 ✓.10✓.35 4800+ / c213.15
expE_sync 🏆I2V-14BscratchLoRA-64 ✓17 Ⓒ .85 ✓.10✓.35 3100+ / run13.20
expF (+ray)I2V-14BscratchLoRA-64 ✓17✓6 Ⓒ .85 ✓.10✓.35 3200+ / run13.05
expF_syncI2V-14BscratchLoRA-64 ✓17✓6 Ⓒ .85 ✓.10✓.35 3100+ / run13.07
expJ_narrowI2V-14BscratchLoRA-64 ✓17 Ⓓ 1−0.2k ✓.10✓.35 1700+ / run
expJ_rayI2V-14BscratchLoRA-64 ✓17✓6 Ⓓ 1−0.2k ✓.10✓.35 1700+ / run
— VACE arm —
expGVACE-14BscratchLoRA-64 + vace-full ✓96 Ⓒ .85 ✓.10 null_vace✓.35 2600+ / relaunch lr1e-513.13@1.5k*
expG0 (clean ctrl)VACE-14BscratchLoRA-64 + vace-full ⚠5e-5 ✓96 ✗ global tnone (τ≡1) ✗ x0 only✗ frozen ✗ full denoise1700+ / run
— 1.3B sandbox (T2V, full-FT) —
expKT2V-1.3Bscratchfull-FT ✓17 Ⓒ .85 ✓(full)✓.10✓.35 5700+ / run12.60@5k
expL_uniT2V-1.3Bscratchfull-FT ✓ ctx→attn ✗ global tnone ✗ x0 only✓(full) ctx-pinned6800+ / run11.08@6k
expL_nkT2V-1.3Bscratchfull-FT ✓ ctx→attn Ⓓ 1−0.2k ✓(full)✓(b2 mix)✓(slice mix) ctx-pinned6800+ / run12.88@6k
— external reference —
CameraAnythingWan T2V-14B ✓(their pluc-adaln)source video none theirs ✗ 50-step stdfrozen release9.75
✎ per-tok AdaLN column = the time-modulation path is trained (--train_modulation, or subsumed by full-FT). * expG score is its pre-crash best @1500; relaunched from s1500 at lr 1e-5. ⚠ expG0 trains the full vace branch at lr 5e-5 — the same config that fried expG; on the @2000–2500 watch list.

Inference-side decisions (apply fleet-wide, not per-experiment)

knobverdictevidence page
denoise steps24 forwards everywhere (audited; only deliberate sweeps differ)24-vs-50
calendarSYNC synchronized-start (τ.85, no lower)expH
trust regionscluster-128tausweep
shift1 for metrics comparability; 5 for viz deliverablesshift sweep
CFGsweep incomplete (3 shards queued); cfg5 costs 2× forwardspending
mem slotremoved everywherereinfer3k
Links: project · atlas (k→τ curves) · parade · external-vs-sandbox board · fleet