MiniT2I-B/32 Canvas-Flow Grid2x2-50 (CC12M pretraining)
Training checkpoints uploaded every 5,000 optimizer steps to checkpoints/checkpoint_<step>.pt.
Each file holds model, ema, optimizer, step and the full training config
(resumable with mini_t2i.train --resume-from).
- 50% normal 512px images, 50% ordered 2x2 canvases (256px cells)
- Normal images: long LLaVA-NeXT captions; grid cells: short LLaVA-NeXT captions
- Exact global shuffle over 10,968,538 CC12M pairs, global batch 1024 on 8 GPUs
- Config:
configs/pretrain_canvas_grid2x2.yml