Upload zimg-fdloss/FINAL_TABLE.md with huggingface_hub

zimg-fdloss/FINAL_TABLE.md

@@ -33,6 +33,13 @@
 | g0c (V1g0 CLEAN rerun) | Base | iMF 3k → FD 2k | iMF(50/25/25) ; SIMcx | data ; rand | 0 | 0.739 | 0.863 | 1.527 | 1.583 |
 | **D1c [CHAMPION] (no r=0 group)** | Base | iMF 3k → FD 2k | iMF(50/0/50) ; SIMcx | data ; rand | 0 | 0.842 | 0.900 | 1.519 | 1.612 |
 | D2c (NEG: iMF-paper lognorm) | Base | iMF 3k → FD 2k | iMF(50/0/50, lognorm) ; SIMcx | data ; rand | 0 | 0.771 | 0.776 | 1.647 | 1.840 |
+| 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 | — | — |
+| 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
 
@@ -42,6 +49,10 @@
 
 **‡ HOLDOUT-LEAK CAVEAT (found 2026-08-02 by code review)**: the original iMF data loader applied the holdout modulus to a shuffled enumeration, so pre-fix iMF-line arms (M, Mg*, V1*, V2–V4, E2, D3 — all rows above g0c that include an iMF stage) trained on ~11/12 of the true ref holdout: their eval-prompt FDr/PickScore carry an optimistic bias (GenEval unaffected — disjoint prompts; turbo-line arms unaffected). g0c/D1c/D2c are post-fix clean reruns; the clean g0c rerun reproduces the old champion within noise @8 (McNemar z=−1.47) and lands lower @1 (z=−4.78) — the V1g0@1 number was partly run-specific. Within-line contrasts among ‡ rows (e.g. D3 vs V1g0) remain internally valid.
 
+**TFC / redistribution law (2026-08-04)**: TFC = tie the 1-jump endpoint's SigLIP2 features to sg(φ(the model's own multi-step rollout endpoint)) — self-distillation of the self-map, externally anchored. Three configs × 2 seeds establish that map-consistency pressure REDISTRIBUTES GenEval between NFE operating points and never creates it: live targets full-run (T1) = hard equalization (@1 record 0.879, z=+3.16; @8 z=−5.53); EMA targets (T2) = inert; champion-init 500-step finishing pass (T3/T3b) = best exchange rate (@1 +2.31z/+1.53z; @8 −1.24z/−2.57z; ~1.2 points gained per point lost). Self-referential signals carry no new information — they shape, external anchors (data, frozen reps) create. TFC weight/horizon = a deployment NFE-preference dial on one recipe.
+
+**Self-Flow probe (arXiv 2603.06507, 2026-08-03)**: generator-self-features fail in BOTH roles. As FD metric: pure self-FD optimizes its own objective (8259→850) while every external signal collapses (GenEval@1 0.835→0.289, z=−23.7 — the init's 1-step capability is destroyed); as a 5th additive term it is a net drag (z=−3.6/−7.8). As iMF self-distillation aux (EMA teacher, 25% token noise asymmetry): rep-cos saturates ≈1.0 by step 500 (signal goes vacuous), GenEval@8 tie (z=−0.83), @1 significantly worse (z=−3.13). Replicates the video program's self-feature negative across modality and σ regimes: the self-encoder's null space contains the generator's own failure modes — external frozen reps are load-bearing. (Cross-validation: Self-Flow's own T2I ablation found uniform > logit-normal timestep dist — matches our D2c rejection.)
+
 **Zhengyang-feedback verdicts (all clean-line, 2026-08-02)**: (1) fewer r=0 samples — CONFIRMED, promoted to champion (D1c); (2) iMF-paper logit-normal(−0.4,1) time dist — REJECTED for z-img@1K (D2c: GenEval 0.776@8 < turbo, FDr worst of the clean line; the shifted-uniform matches z-img's high-noise training emphasis); (3) dropping Inception — REJECTED (D3: GenEval tie z=−0.42/−0.23, Pick/CLIP mildly worse; 2nd confirmation of round-1 SIM>SM).
 
 NEG = registered negative results kept for the record (SIMu: one-joint entanglement; V2: joint-from-scratch; E1/E2: rMMD queue-staleness; D2c: lognorm time dist). Caveats: BLIP3o contains geneval_train (paired compositional supervision iMF exploits, turbo never saw); VLM-judge GenEval anchor 0.806 vs ~0.83 official detector; single seed per arm; 5k total steps is undertrained-scale.