A short decoder-only finetune, run after the end-to-end stage. It is optional and cheap — 10k steps against 600k — and worth about 0.2 unguided gFID on ImageNet-256 (3.00 → 2.79 at 80 epochs). Every ImageNet-256 number reported for this repo uses it.
INIT_CKPT=exps/e2e-f8d4-sitxl2-256-phase2/checkpoints/0600000.pt \
bash scripts/train_decoder_ft.sh| trainable | vae.decoder, vae.post_quant_conv, discriminator |
| frozen | vae.encoder, vae.quant_conv |
| loss | L1 + LPIPS + GAN — no SiT, no REPA alignment |
| steps | 10k, at global batch 256 |
| resolution | 256 only |