Upload metrics/METRICS_25K_SNAPSHOT.md with huggingface_hub

metrics/METRICS_25K_SNAPSHOT.md

@@ -1,8 +1,8 @@
 # DeCo GAN-Finetune — 25k-step Snapshot
 
 All finetunes compared at the **same 25 000-step checkpoint**; `original` (pretrained DeCo, no finetune)
-is the reference. Methods that have **not yet reached 25k are excluded** for now: `gan_dinov2_t035`
-(only to 20k) and `drift` (only to 15k) — add once they hit 25k.
+is the reference. `gan_dinov2_t035` now reaches 25k and is included (full suite). Only `drift` remains
+excluded (its single ckpt is at 15k).
 
 Arrows: ↑ higher-better, ↓ lower-better. `—` = not measured for that model. `n=10000` for COCO & quality.
 
@@ -12,25 +12,28 @@ Arrows: ↑ higher-better, ↓ lower-better. `—` = not measured for that model
 | baseline               | 0.8160 | 0.8564 | 36.818 | 36.982 | 0.8097 | 0.4156 | 0.7062 | 75.464 | 0.6320 | 29.162 |
 | gan_patchgan           | 0.8166 | 0.8581 | 37.103 | 37.029 | 0.8149 | 0.3961 | 0.7017 | 75.422 | 0.6270 | 29.288 |
 | gan_patchgan_bn        | 0.8169 | 0.8570 | — | — | — | — | — | — | — | — |
-| gan_dinov2             | 0.8285 | 0.8567 | 34.887 | **38.989** | 0.7898 | **0.5105** | **0.7803** | 76.373 | 0.7149 | — |
-| gan_dinov2_50          | 0.8306 | 0.8576 | 34.817 | 36.807 | 0.7805 | 0.4745 | 0.7750 | **76.640** | **0.7217** | 27.319 |
+| gan_dinov2             | 0.8285 | 0.8567 | 34.887 | **38.989** | 0.7898 | **0.5105** | 0.7803 | 76.373 | 0.7149 | — |
+| gan_dinov2_50          | 0.8306 | 0.8576 | 34.817 | 36.807 | 0.7805 | 0.4745 | 0.7750 | **76.640** | 0.7217 | 27.319 |
 | gan_dinov2_all         | 0.8261 | 0.8542 | — | — | — | — | — | — | — | — |
-| **gan_dinov2_text**    | **0.8327** | 0.8515 | **32.476** | 37.202 | **0.7246** | 0.4784 | 0.7666 | 76.489 | 0.7100 | **25.526** |
+| **gan_dinov2_t035**    | **0.8337** | 0.8575 | 34.324 | 37.978 | 0.8092 | 0.4976 | **0.7827** | 76.577 | **0.7263** | 27.661 |
+| gan_dinov2_text        | 0.8327 | 0.8515 | **32.476** | 37.202 | **0.7246** | 0.4784 | 0.7666 | 76.489 | 0.7100 | **25.526** |
 | gan_dinov2_50_longcap  | 0.8144 | 0.7113 | — | — | — | — | — | — | — | — |
 
 **Notes**
 - `original` has no "25k" (it is the base model); its own values are used as the reference row. Its DPG
   (~0.8158) was measured on the pretrained checkpoint (not stored per-step on HF).
-- COCO-10K measured for: original, baseline, gan_patchgan, gan_dinov2, gan_dinov2_50, gan_dinov2_text (@25k) — others `—`.
+- COCO-10K measured for: original, baseline, gan_patchgan, gan_dinov2, gan_dinov2_50, gan_dinov2_t035, gan_dinov2_text (@25k) — others `—`.
 - Visual quality measured for: original, baseline, gan_patchgan, gan_dinov2 (NR-IQA only), gan_dinov2_50,
-  gan_dinov2_text — `gan_dinov2` pFID still missing.
+  gan_dinov2_t035, gan_dinov2_text — `gan_dinov2` pFID still missing.
 
 **Read at a glance (25k)**
-- **DPG**: DINOv2-disc GANs win; `gan_dinov2_text` best (0.8327), then dinov2_50 (0.8306) > dinov2 (0.8285)
-  > dinov2_all (0.8261). baseline/patchgan/patchgan_bn ~0.816; longcap collapsing (0.8144).
+- **DPG**: DINOv2-disc GANs win; `gan_dinov2_t035` best (0.8337), then gan_dinov2_text (0.8327) > dinov2_50
+  (0.8306) > dinov2 (0.8285) > dinov2_all (0.8261). baseline/patchgan/patchgan_bn ~0.816; longcap 0.8144.
 - **GenEval**: all ~0.85–0.86 (GAN ≈ baseline; realism, not composition); longcap crashed to 0.7113.
-- **COCO realism**: DINOv2 GANs best FID/IS/Recall; patchgan worst FID (37.1, above original).
-- **Visual quality**: original ≪ baseline ≈ patchgan ≪ DINOv2 GANs; best pFID = gan_dinov2_text (25.53).
+- **COCO realism**: `gan_dinov2_text` best FID (32.5) + CMMD (0.725); `gan_dinov2` best IS/Recall; patchgan
+  worst FID. Note `t035` has strong FID (34.3) but **worst CMMD (0.809, ≈baseline)** despite top DPG.
+- **Visual quality**: original ≪ baseline ≈ patchgan ≪ DINOv2 GANs. `t035` best NR-IQA (TOPIQ 0.783 / MANIQA
+  0.726); `gan_dinov2_text` best pFID (25.53).
 
 ## P-IDS / U-IDS @25k (CoModGAN realism; higher-better, max 0.5; side-signal only — very noisy near floor)
 
@@ -41,8 +44,9 @@ Arrows: ↑ higher-better, ↓ lower-better. `—` = not measured for that model
 | gan_patchgan | 0.0077 | 0.0002 |
 | **gan_dinov2** | **0.0235** | **0.0030** |
 | gan_dinov2_text | 0.0152 | 0.0012 |
+| gan_dinov2_t035 | 0.0083 | 0.0001 |
 | gan_dinov2_50 | 0.0038 | 0.0000 |
 
 gan_dinov2 most realistic here (≈3× others), then gan_dinov2_text; absolute values tiny (SVM still
 separates real vs gen ~98%+). Precision (dropped from main table) @25k: patchgan 0.407 > baseline 0.402 >
-original 0.380 > gan_dinov2 0.351 > dinov2_50 0.323 > drift(15k) 0.301 — inverse of realism (P/R tradeoff).
+original 0.380 > gan_dinov2 0.351 > t035 0.334 > dinov2_50 0.323 > drift(15k) 0.301 — inverse of realism (P/R tradeoff).