Zaytron40k/2511-style-transfer-dual

t01nstyle Edit-LoRA for Qwen-Image-Edit-2511 (dual-captioning)

Edit-mode style-transfer LoRA for Qwen-Image-Edit-2511. Trained with control conditioning (sketches + photos) and dual-captioning (detailed + trigger-only) for clean style/content separation.

Status: training complete (15 epochs).

Trigger: t01nstyle

Recommended checkpoint

Start with t01nstyle_edit2511_dual-000013.safetensors (global minimum loss at epoch 13: 0.04599). A/B test against epochs ±2 by visual quality.

Per-epoch loss

epoch	loss
1	0.05171
2	0.05004
3	0.04940
4	0.04978
5	0.04802
6	0.04946
7	0.04896
8	0.04922
9	0.04740
10	0.04852
11	0.04667
12	0.04619
13	0.04599 (min)
14	0.04651

Inference

• Pipeline: Qwen-Image-Edit-2511 with control image (sketch or photo)
• LoRA scale: 1.0 to start, adjust by visual quality
• Sampler: FlowMatch / Euler, 25-40 steps, true CFG 4.0
• Resolution: target ~1024^2, control 512^2 (training resolution)

Training

• Tuner: musubi-tuner (kohya-ss), --model_version edit-2511 --edit_plus
• Captioning: dual (1:1 mix of detailed captions + trigger-only)
• Dataset: 256 art targets x 2 control sets (sketches + photos) x 2 captioning modes = 1024 effective pairs
• Network dim/alpha: 32/32 with LoRA+ (loraplus_lr_ratio=2)
• Target modules: 840 (60 transformer blocks x 14 - full attention + MLP + img_mod.1 + txt_mod.1 via include_patterns rescue)
• Learning rate: 5e-5 (adamw8bit, constant_with_warmup, 200 warmup steps)
• Timestep sampling: qwen_shift
• Batch size: 4
• 15 epochs, ~5010 steps, control_resolution 512^2
• Hardware: H100 SXM 80GB (bf16, sdpa)

Files

• t01nstyle_edit2511_dual-NNNNNN.safetensors - checkpoints per epoch
• train.sh - full training command
• dataset.toml - 4-block dataset configuration
• training.log - full training log
• tensorboard/ - tensorboard event files for loss curves