Revert to last working version (verbose logs + VAE + simple MSE loss — before v3/v4 broke training)

Revert to v4 epoch-based loop + fix streaming repeat + fix scheduler + remove data std norm

v4.1: Fix training loop — step-based (not epochs), auto-cycle dataset, fix streaming exhaust, fix scheduler warning

v4: Add Min-SNR-γ + velocity direction loss + CCA + multi-scale loss + gate bias (DeepSeek/FasterDiT/DiCo/DiMR research)

v3: Add large anime/art datasets, cosine-with-restarts schedule, 2x LR, smart warmup, streaming support, resume training

Add verbose training logs: ETA, loss trend, speed, VRAM, grad norm, epoch summaries

Optimize notebook: 40% faster CfC blocks, simplified spatial mix

v2: Add VAE latent training, fix datasets, streaming support

V2: Add VAE, fix datasets, streaming, precaching

Upload LiquidDiffusion_Training.ipynb

Add Colab training notebook