BiliSakura/RSEdit-DiT-experimental

RSEdit DiT Modified (NoWindow RoPE + QKNorm + SEG)

This directory stores a local Diffusers-compatible checkpoint for the RSEdit DiT modified architecture experiment (NoWindow RoPE, QKNorm, and SEG variant).

Contents

• model_index.json: pipeline/component registry (RSEditModifiedDiTPipeline).
• pipeline.py: custom local pipeline implementation used for inference/loading.
• checkpoint-30000/: training-state snapshot at step 30k (optimizer/scheduler/random state plus transformer weights).
• transformer/, vae/, text_encoder/, tokenizer/, scheduler/: exported model components for direct pipeline loading.
• logs/: experiment logs (logs/rsedit-dit/...).

Quick Load (Diffusers)

from diffusers import DiffusionPipeline

model_dir = "datasets/my_running_checkpoints/RSEdit-DiT-Modified-NoWindow-RoPE-QKNorm-SEG"
pipe = DiffusionPipeline.from_pretrained(model_dir, trust_remote_code=True)
pipe = pipe.to("cuda")

Notes

• checkpoint-30000/optimizer.bin is large and mainly needed for training resume, not standard inference.
• checkpoint-30000/transformer_ema/ exists for EMA tracking during training workflows.
• Keep pipeline.py and model_index.json aligned if the custom pipeline class changes.