Zaytron40k/t01nstyle-edit2511

t01nstyle LoRA for Qwen-Image-Edit-2511

LoRA for Qwen-Image-Edit-2511 trained with musubi-tuner to transform sketches/photos into the target style t01nstyle.

Trigger / caption format

Transform into t01nstyle. <description of content>

Dataset

Three synchronized sets with matching basenames:

• TargetStyleArt — target/style images + captions
• Control_sletch_deformed — sketch controls
• controlPhoto — photo controls

This run uses the accepted softer crop variant where former square images were cropped to about AR 1.13.

Training summary

• Tuner: musubi-tuner
• Model version: edit-2511
• Network dim / alpha: 32 / 32
• Optimizer: adamw8bit
• Learning rate: 1e-4
• Scheduler: constant_with_warmup
• Warmup steps: 200
• Timestep sampling: shift
• Discrete flow shift: 2.6
• Precision: bf16
• Epochs: 40
• Checkpoint save interval: every 2 epochs

target_modules

The current musubi default excluded modulation layers, so this run was started with an explicit exact 12-module 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

Files

• checkpoints/ — ready checkpoints uploaded so far
• config/dataset.toml — active dataset config
• config/train_t01nstyle_edit2511_exact12.sh — active train command/script