Zaytron40k/2511-surgical-v3-t01nstyle

t01nstyle LoRA for Qwen-Image-Edit-2511 (surgical v3)

Style LoRA trained in T2I mode (--model_version original) on Edit-2511 weights with image-stream-only surgical targeting (image-stream AdaLN modulation + image-stream MLP down-projection). Designed to preserve native multi-reference editing and InstantX ControlNet compatibility.

Status: training complete (22 epochs).

Trigger: t01nstyle

Recommended checkpoint

Start with t01nstyle_qie2511_t2i_surgical_v3-000009.safetensors (global minimum loss at epoch 9: 0.07330). A/B test against epochs +/- 2 by visual quality.

Per-epoch loss

epoch	loss
1	0.07905
2	0.07744
3	0.07633
4	0.07643
5	0.07635
6	0.07627
7	0.07951
8	0.07611
9	0.07330 (min)
10	0.07634
11	0.07681
12	0.07656
13	0.07567
14	0.07650
15	0.07673
16	0.07683
17	0.07779
18	0.07532
19	0.07647
20	0.07490
21	0.07601
22	0.07713

Loss curve typically very flat for surgical LoRA (small parameter count). Visual quality at lora_scale 1.0 is the deciding factor.

Targeting (v3)

Two patterns (vs three in v1):

• img_mod.1 - image-stream AdaLN modulation
• img_mlp.net.2 - image-stream MLP down-projection

120 LoRA modules total (60 transformer blocks x 2). All attention projections (to_q/k/v, to_out.0, add_q/k/v_proj, to_add_out) are bit-identical to base Edit-2511. MLP up-projection (img_mlp.net.0.proj) and entire text-stream untouched.

Inference

• Pipeline: Qwen-Image-Edit-2511 (multi-reference)
• Compatible with InstantX ControlNet for Qwen-Image (use controlnet_conditioning_scale 0.6-0.8)
• LoRA scale: 1.0 (surgical LoRA - full strength expected)
• Sampler: FlowMatch / Euler, 25-40 steps, true CFG 4.0
• Resolution: 1024^2 or native 1328^2

Training

• Tuner: musubi-tuner (kohya-ss), --model_version original on Edit-2511 weights
• Dataset: 256 art images in t01nstyle, T2I captions (no control images)
• Network: rank 16 / alpha 16
• Learning rate: 5e-5 (adamw8bit, constant_with_warmup, 200 warmup steps)
• Timestep sampling: shift, discrete_flow_shift 2.2, num_timestep_buckets 8
• 22 epochs, ~3036 steps
• Hardware: H100 SXM 80GB (bf16, sdpa)

The unnumbered t01nstyle_qie2511_t2i_surgical_v3_final.safetensors corresponds to epoch 22.

Files

• t01nstyle_qie2511_t2i_surgical_v3-NNNNNN.safetensors - checkpoints per epoch
• train.sh - full training command
• dataset.toml - dataset configuration
• training.log - full training log
• tensorboard/ - tensorboard event files for loss curves