Zaytron40k/t01nstyle-qie2511-edit-40ep

t01nstyle Qwen-Image-Edit-2511 LoRA (40-epoch run)

LoRA for Qwen-Image-Edit-2511 trained with musubi-tuner for paired edit-mode sketch → stylized image generation in style t01nstyle.

Caption / trigger format

t01nstyle, <description>

Dataset

Paired dataset with matching basenames:

• TargetStyleArt — target/style images + captions
• ControlSketch — sketch controls

Training setup

• model version: edit-2511
• network dim / alpha: 32 / 32
• optimizer: adamw8bit
• learning rate: 1e-4
• scheduler: constant_with_warmup
• warmup steps: 400
• timestep sampling: shift
• discrete flow shift: 2.2
• timestep buckets: 4
• epochs: 40
• checkpoint save interval: every 1 epoch
• batch size: 2
• seed: 42

Exact target modules

This run uses an explicit exact 12-pattern target selection:

• to_q
• to_k
• to_v
• add_q_proj
• add_k_proj
• add_v_proj
• to_out.0
• to_add_out
• img_mlp.net.2
• img_mod.1
• txt_mlp.net.2
• txt_mod.1

Observed module count at launch: 720 LoRA modules.

Files

• checkpoints/ — uploaded epoch checkpoints for this run
• config/dataset.toml — active dataset config
• config/train_qwen2511_edit_runpod.sh — launched train script
• logs/training.log — current training log snapshot