--- license: apache-2.0 base_model: Qwen/Qwen-Image-Edit-2511 tags: - lora - qwen-image-edit - qwen-image-edit-2511 - style - musubi-tuner --- # t01nstyle LoRA for Qwen-Image-Edit-2511 Style LoRA trained in T2I mode (`--model_version original`) on Edit-2511 weights with surgical Tier B+ targeting (image-stream AdaLN modulation + image-stream MLP only). Designed to preserve native multi-reference editing and InstantX ControlNet compatibility. **Status:** training complete (25 epochs, 3450 steps). **Trigger:** `t01nstyle` ## Recommended checkpoint Start with `t01nstyle_qie2511_t2i_surgical-000018.safetensors` — global minimum loss + center of the sweet spot. A/B test plan (LoRA scale 1.0): 1. **epoch 18** — primary candidate (loss 0.07351, the minimum) 2. **epoch 16** — if 18 looks too strong / distorts the base 3. **epoch 22** — if 18 looks too weak in style The final unnumbered checkpoint `t01nstyle_qie2511_t2i_surgical_final.safetensors` corresponds to epoch 25. ## Per-epoch loss | epoch | loss | | epoch | loss | |------:|:--------|--|------:|:--------| | 1 | 0.07772 | | 14 | 0.07651 | | 2 | 0.07733 | | 15 | 0.07539 | | 3 | 0.07672 | | **16**| **0.07353** | | 4 | 0.07606 | | 17 | 0.07477 | | 5 | 0.07701 | | **18**| **0.07351** (min) | | 6 | 0.07731 | | 19 | 0.07443 | | 7 | 0.07579 | | 20 | 0.07462 | | 8 | 0.07678 | | 21 | 0.07602 | | 9 | 0.07644 | | 22 | 0.07664 | | 10 | 0.07666 | | 23 | 0.07661 | | 11 | 0.07833 | | 24 | 0.07404 | | 12 | 0.07759 | | 25 | 0.07389 | | 13 | 0.07379 | | | | Loss curve is very flat (range 0.07351–0.07833, ~6%) — typical for surgical LoRA with low parameter count. Visual quality at lora_scale 1.0 is the deciding factor. ## 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: 10242 or native 13282 ## 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, surgical Tier B+ targeting via `--network_args` - Target modules: `img_mod.1`, `img_mlp.net.0.proj`, `img_mlp.net.2` (image stream only, attention untouched) - 180 LoRA modules total (60 transformer blocks × 3) - Learning rate: 1e-4 (adamw8bit, constant_with_warmup, 200 warmup steps) - Timestep sampling: shift, discrete_flow_shift 2.2, num_timestep_buckets 4 - 25 epochs, 3450 steps, ~2h 00m wall time - Hardware: H100 SXM 80GB (bf16, sdpa attention) **Why this targeting:** 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. Multi-reference image embeddings flow through joint attention unchanged. ControlNet residuals land on unchanged hidden states. The LoRA modulates only per-block style injection (AdaLN) and image-stream FFN output. ## Files - `t01nstyle_qie2511_t2i_surgical-000001.safetensors` ... `-000024.safetensors` — checkpoints per epoch (1–24) - `t01nstyle_qie2511_t2i_surgical_final.safetensors` — epoch 25 final - `train.sh` — exact training command - `dataset.toml` — dataset configuration - `training.log` — full training log - `tensorboard/` — tensorboard event files for loss curves
