update model card: best quick_fid_4096=104.922272 step=50000

README.md

@@ -9,7 +9,7 @@ tags:
 - jepa
 ---
 
-# Soft-JEPA-Flow (jepa_full_final)
+# Soft-JEPA-Flow (jepa2_run)
 
 This repository stores checkpoints automatically uploaded from training.
 
@@ -22,27 +22,24 @@ This repository stores checkpoints automatically uploaded from training.
 - MLP ratio: `4.0`
 - Latent shape: `32x32x4`
 - Number of classes: `100`
-- Student layer: `3`
-- Teacher layer: `7`
-- Mask ratio: `0.3`
-- Lambda JEPA: `0.15`
+- JEPA branch: disabled in baseline mode.
 
 ## Training Objective
-- Mode: `jepa`
-- Objective: L_total = L_gen + lambda_jepa * L_jepa (teacher-student JEPA objective with EMA teacher).
-- Optimizer: `adamw` (lr=`0.0001`, beta1=`0.9`, beta2=`0.99`, weight_decay=`0.02`)
-- Timestep schedule: `lognormal` (mean=`-0.4`, std=`1.0`)
+- Mode: `jepa2`
+- Objective: L_total = L_gen (rectified-flow velocity prediction objective).
+- Optimizer: `adam` (lr=`0.0001`, beta1=`0.9`, beta2=`0.99`, weight_decay=`0.04`)
+- Timestep schedule: `uniform` (mean=`-0.4`, std=`1.0`)
 
 ## Metrics
 - Best metric key: `quick_fid_4096`
-- Best metric value: `48.418213`
-- Best checkpoint step: `150000`
+- Best metric value: `104.922272`
+- Best checkpoint step: `50000`
 
 ## Upload Timing (Automatic)
 - Upload policy: upload whenever a **new best metric** is observed.
-- Upload timestamp (UTC): `20260306-013603-UTC`
-- Run name: `jepa_full_final`
-- Path for this artifact: `jepa_full_final/best/step_150000_20260306-013603-UTC`
+- Upload timestamp (UTC): `20260309-213746-UTC`
+- Run name: `jepa2_run`
+- Path for this artifact: `jepa2_run/best/step_50000_20260309-213746-UTC`
 
 ## Evaluation Notes
 - FID pipeline uses latent decode with `stabilityai/sd-vae-ft-mse`, then InceptionV3 features.