t01nstyle-qie2509-lora

t01nstyle LoRA for Qwen-Image-Edit-2509

Style LoRA trained in T2I mode (--model_version original) on Qwen-Image-Edit-2509 weights. Loads on the full Qwen-Image-Edit-2509 pipeline at inference.

Trigger: t01nstyle

Recommended inference setup:

• Pipeline: Qwen-Image-Edit-2509 (multi-reference + ControlNet pose stack)
• LoRA scale: 0.7–0.9 (not 1.0 — preserves identity from reference images)
• Sampler: FlowMatch / Euler, 25–40 steps
• True CFG: 4.0
• Resolution: 1024² or native 1328²

Training:

• Tuner: musubi-tuner (kohya-ss), --model_version original
• Dataset: 256 art targets in t01nstyle, captions only (no control images)
• Network rank/alpha: 32 / 32
• Learning rate: 1e-4 (adamw8bit, constant_with_warmup, 200 warmup steps)
• Timestep sampling: shift, discrete_flow_shift 2.2, num_timestep_buckets 4
• 30 epochs, ~4,140 steps
• Hardware: H100 SXM 80GB

Why T2I mode on an Edit model: training in T2I mode (without control images) gives the LoRA a clean style-only signal. At inference on the Edit pipeline, ControlNet handles pose, multi-reference handles identity, and this LoRA only handles style — no conflicts between components.

Checkpoints provided: epochs 12–28 (every 2 epochs) + final (epoch 30). Best checkpoint typically around epochs 18–25.