Skip to content

Latest commit

 

History

History
21 lines (14 loc) · 900 Bytes

DECODER.md

File metadata and controls

21 lines (14 loc) · 900 Bytes

Decoder finetune

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

What it trains

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