Upload zimg-fdloss/RECIPE.md with huggingface_hub

zimg-fdloss/RECIPE.md

@@ -0,0 +1,70 @@
+# Flow Matching Mid/Post-training — Final Recipe Card (frozen 2026-08-04)
+
+Model: Z-Image 6B S3-DiT @1024². Reference data: BLIP3o-60k (GPT-4o set),
+holdout = sorted-enumeration idx % 12 == 0.
+
+## The recipe
+
+1. **Mid-training (coverage-preserving transport)** — iMF/MeanFlow on the
+   base model, 3k steps, lr 2e-5: boundary parameterization u(z,t,t)=v via
+   zero-init r-branch; v-compound loss with central-difference dF (h=0.01);
+   (s,r) = max/min of two U(0,1) + static shift 3.0; groups 50% r=s (FM
+   anchor) / 0% r=0 / 50% spans — **no explicit full-jump group, no guidance
+   fusion (g=0)**. Mode-seeking is deferred entirely to stage 2.
+2. **Post-training (externally anchored conditional distribution matching)**
+   — coupled FD, 2k steps, lr 1e-5 cosine: random-NFE {1,2,4,8} rollout with
+   one gradient step + direct endpoint jump; loss = Eq.6-normalized sum of
+   Gaussian FD terms in frozen SigLIP2-joint / Inception / MAE spaces + a
+   standalone cross-covariance term on a factor-probe basis; conditional
+   coupling via top-32 diag-whitened CCA text coordinates (K=32 optimal;
+   K=64 does not transfer); EMA population moments (β=0.999). **Global
+   batch 64 (4-node) if available** — the sole surviving upgrade.
+3. **Optional TFC finishing dial** — short (250–500 step) pass tying the
+   1-jump endpoint's SigLIP2 features to sg(own multi-step endpoint):
+   redistributes quality between NFE operating points along a measured
+   exchange curve (~1.2 pts gained @1 per pt lost @8). Not net creation.
+
+## Headline numbers (GenEval macro, 553×2 VLM-judged; FDr held-out judge mean)
+
+| model | GenEval@8 | GenEval@1 | FDr@8 | FDr@1 |
+|---|---|---|---|---|
+| Z-Image-Turbo (distilled baseline) | 0.811 | 0.719 | 2.174 | 1.700 |
+| **champion (this recipe, b16)** | **0.900** | **0.842** | 1.612 | 1.519 |
+| champion @ batch 64 | 0.910 | 0.846 | 1.621 | 1.447 |
+| 1-NFE specialist (TFC full-run) | 0.843 | 0.879 | 1.909 | 1.596 |
+
+McNemar vs turbo: z=+7.0/+8.2 (Holm-robust). No teacher, no discriminator,
+no fake-score network, no CFG at train or inference; ~4.5 h on one 8×H200
+node (b16).
+
+## The three laws (measured)
+
+1. **Division of labor**: mid-training must not spend mode coverage
+   (g=0 beats g∈{2,4}; removing the r=0 group is the largest single gain;
+   every stage-mixing variant failed). Coverage is a training resource the
+   FD stage converts into endpoint quality.
+2. **Redistribution law**: self-referential signals (self-features,
+   self-distillation, self-map consistency) redistribute quality across NFE
+   operating points but do not create it — ~20 arms, zero replicated
+   counterexamples. Creation tracks external information only.
+3. **Information economics**: extractable population signal beyond two
+   moments in the coupled feature space is ~1% of the FM loss and largely
+   moment-redundant. Kernel drift, learned score-differences (including
+   evidence-gated), adversarial heads, and larger coupling bases all fail
+   to beat second-moment FD + EMA. The Gaussian approximation's fast
+   saturation acts as an implicit regularizer protecting the conditional
+   coupling terms' gradient share.
+
+## Checkpoints in this folder
+
+- `champion-D1c-clean/` — the b16 champion (FD ckpt + iMF init)
+- `champion-D1c-b64/` — batch-64 variant (best both-points nominal)
+- `T1-1nfe-record/` — best pure-1-NFE model (0.879@1)
+- `T3-balanced-finisher/`, `T4a-near-pareto-finisher/` — TFC dial points
+- `V1g0-clean-rerun-g0c/`, `champion-V1g0/` — recipe-ancestor reference
+- `reviewer-SIMv2/`, `reviewer-SMv2/`, `turbo-line-SIMcx/`, `V1prog-g4/`,
+  `V3-bestFDr8/` — turbo-line and ablation references
+
+Full report: Notion "Flow Matching Mid/Post-training" (rev 5) — 40-arm
+codename map, 25-row registered-prediction ledger, estimator matrix,
+implementation-integrity audits. sha256 of every checkpoint: manifest.tsv.