recon-ckpt

Checkpoints from the RECON long-run study (SiT-XL/2 + REPA + one pooled DINOv2-B semantic token, ImageNet-256, SD-VAE f8d4). All files are full training checkpoints (model, ema, opt, args, steps, epoch); evaluate the ema weights. FID below: unguided, 50k samples, SDE Euler-Maruyama 250 steps, ADM reference (VIRTUAL_imagenet256_labeled.npz).

bs-1024 no-CFG run (recon-nocfg-bs1024-*)

Recipe: batch 1024, AdamW lr 2e-4 betas (0.9, 0.95), wd 0, EMA 0.9995, no class dropout (1000-row label table; guide with autoguidance, not CFG), token injected for t <= 0.5 in training, proj-coeff 1.5, FlexAttention. const-*: constant lr 2e-4 (epochs 0-320). lrdecayA-*: continued from const-ep0320 with a linear lr decay 2e-4 -> 2e-5 over steps 400,320 -> 1,000,800 (epochs 320-800). 1,251 steps per epoch.

file	epoch	step	FID@250
recon-nocfg-bs1024-const-ep0080.pt	80	100,080	3.43
recon-nocfg-bs1024-const-ep0160.pt	160	200,160	2.77
recon-nocfg-bs1024-const-ep0240.pt	240	300,240	2.50
recon-nocfg-bs1024-const-ep0320.pt	320	400,320	2.40
recon-nocfg-bs1024-lrdecayA-ep0380.pt	380	475,380	2.49 (superseded; kept for reference)
recon-nocfg-bs1024-lrdecayA-ep0400.pt	400	500,400	2.27
recon-nocfg-bs1024-lrdecayA-ep0480.pt	480	600,480	2.21
recon-nocfg-bs1024-lrdecayA-ep0560.pt	560	700,560	2.16
recon-nocfg-bs1024-lrdecayA-ep0640.pt	640	800,640	2.11
recon-nocfg-bs1024-lrdecayA-ep0720.pt	720	900,720	2.14
recon-nocfg-bs1024-lrdecayA-ep0800.pt	800	1,000,800	2.11 (final; reference bs-256 recipe A at 800 ep: 2.21)

Recipe-A (bs 256) intermediate checkpoints

Continuations of SaeedRazavi/RECON recon10ep.ckpt / recon20ep.ckpt for 5 epochs each, original recipe unchanged.

file	epoch	step	FID@250
recon-ep0015.pt	15	75,060	13.69
recon-ep0025.pt	25	125,100	7.06

Code: llm-robotics/RECON-Dev, branch long-run/experiments.