Upload zimg-fdloss/FINAL_TABLE.md with huggingface_hub

zimg-fdloss/FINAL_TABLE.md

@@ -1,6 +1,6 @@
 # FINAL TABLE — FD/iMF post-training of Z-Image 6B @1024², GPT-4o reference (BLIP3o-60k)
 
-**Shared hyperparameters** — FD stages: AdamW(0.9, 0.95) wd 0, lr 1e-5 (warmup 100 + cosine), global batch 16 (8×H200, bs/rank 2), EMA population stats β=0.999, Eq.6 self-norm c=0.01, shrink 1e-4, ckpts @2k steps. iMF stages: lr 2e-5 (warmup 200, const), central-diff dF (h=0.01), (t,r) = max/min of two U(0,1) + shift-3.0, groups 50% r=t / 25% r=0 / 25% span (champion D1c: 50/0/50 — no explicit r=0 group), boundary-v tangent, data = 54k BLIP3o ref pairs (holdout excluded; see ‡ leak caveat below). All sampling CFG-free on the shift-3.0 grid (u-jumps for meanflow models, Euler otherwise). Eval: GenEval 553×2 seeds VLM-judged (identical protocol all rows), FDr = held-out judge mean (dinov2/CLIP/ConvNeXtV2) vs GPT-4o ref at matched n (1000 @8NFE / 256 @1NFE) — FDr comparable within a column only. Prompt pools: v2 category-balanced.
+**Shared hyperparameters** — FD stages: AdamW(0.9, 0.95) wd 0, lr 1e-5 (warmup 100 + cosine), global batch 16 (8×H200, bs/rank 2), EMA population stats β=0.999, Eq.6 self-norm c=0.01, shrink 1e-4, ckpts @2k steps. iMF stages: lr 2e-5 (warmup 200, const), central-diff dF (h=0.01), (t,r) = max/min of two U(0,1) + shift-3.0, groups 50% r=t / 25% r=0 / 25% span (champion D1c: 50/0/50 — no explicit r=0 group), boundary-v tangent, data = 54k BLIP3o ref pairs (holdout excluded; see ‡ leak caveat below). All sampling CFG-free on the shift-3.0 grid (u-jumps for meanflow models, Euler otherwise). Eval: GenEval 553×2 seeds VLM-judged (identical protocol all rows), FDr = held-out judge mean (dinov2/CLIP/ConvNeXtV2) vs GPT-4o ref at matched n (1000 @8NFE / 256 @1NFE) — FDr comparable within a column only. Prompt pools: v2 category-balanced. GenEval numbers are MACRO over the 6 tags; McNemar z-tests are item-level (553x2 pairs; cluster-robust z 2-10% smaller, no claim flips). VLM judge uses exact-count criteria on all tags (stricter than official at-least include semantics); the ~0.83 official anchor is an external report, not reproduced locally. FDr finite-sample sigma ~0.03 (n=1000) / ~0.06 (n=256) on the judge mean - treat sub-0.05 gaps as ties. Family-wise note: headline claims (champion promotion, D1c>turbo, T1 both directions, Self-Flow negatives) survive Holm correction at m=36; ~2-sigma claims are flagged nominal in the journal.
 
 | model | init | stages (steps) | loss terms | rollout | fuse-g | GenEval @1 | GenEval @8 | FDr↓ @1 | FDr↓ @8 |
 |---|---|---|---|---|---|---|---|---|---|
@@ -36,9 +36,9 @@
 | S (NEG: Self-Flow probe, self-FD) | Base | iMF 3k → FD 2k | self-features ONLY (σ.1,L20) | data ; rand | 0 | 0.299 | 0.709 | 3.242 | 3.206 |
 | SIMcxS (NEG: champion + self term) | Base | iMF 3k → FD 2k | SIMcx + self (5th term) | data ; rand | 0 | 0.742 | 0.867 | 1.781 | 1.925 |
 | SD (NEG: iMF self-distill aux) | Base | iMF+selfdist 3k → FD 2k | iMF+L_rep(γ.5) ; SIMcx | data ; rand | 0 | 0.803 | 0.892 | 1.497 | 1.708 |
-| T1 (TFC live, full-run) [1-NFE record] | Base | iMF 3k → FD+TFC 2k | SIMcx + self-map distill(w1) | data ; rand | 0 | 0.879 | 0.843 | — | — |
-| T2 (TFC EMA targets — inert) | Base | iMF 3k → FD+TFC 2k | SIMcx + TFC(EMA.999) | data ; rand | 0 | 0.836 | 0.882 | — | — |
-| T3 (TFC finishing pass, seed1) | Base | …D1c → FD+TFC 500 | SIMcx + TFC(w.5, lr3e-6) | data ; rand | 0 | 0.860 | 0.891 | — | — |
+| T1 (TFC live, full-run) [1-NFE record] | Base | iMF 3k → FD+TFC 2k | SIMcx + self-map distill(w1) | data ; rand | 0 | 0.879 | 0.843 | 1.596 | 1.909 |
+| T2 (TFC EMA targets — inert) | Base | iMF 3k → FD+TFC 2k | SIMcx + TFC(EMA.999) | data ; rand | 0 | 0.836 | 0.882 | 1.561 | 1.651 |
+| T3 (TFC finishing pass, seed1) | Base | …D1c → FD+TFC 500 | SIMcx + TFC(w.5, lr3e-6) | data ; rand | 0 | 0.860 | 0.891 | 1.545 | 1.630 |
 | T3b (finishing pass, seed2) | Base | …D1c → FD+TFC 500 | SIMcx + TFC(w.5, lr3e-6) | data ; rand | 0 | 0.857 | 0.881 | — | — |
 
 ## Champion vs baselines