Downloads · 30 days
0
2511-style-transfer-semireal-v5
Zaytron40k/2511-style-transfer-semireal-v5
2511-style-transfer-semireal-v5 is a machine learning model from Zaytron40k. Use it for the machine learning task on the model card, and read the license before you ship it in a product. The card lists the license as apache-2.0.
Edit-mode style-transfer LoRA. Successor to 2511-style-transfer-semireal (v4) — same hyperparameters, new dataset: 204 target arts + 204 matching semireal-sketch controls + AI-generated detailed captions (Grok-4.20 vi…
Access
Public
Updated Jun 3, 2026
Repo size
14.2 GB
Likes
0
Public
Repo makeup
Click a slice to open those files.
.safetensors8.9 GB · 100%
At a glance
- License
- apache-2.0
- Access
- Public
- Created
- Jun 3, 2026
- Updated
- Jun 3, 2026
- SHA
- 0ed3a94c
Base models
Model card
From the Hugging Face model README
t01nstyle Edit-LoRA v5 for Qwen-Image-Edit-2511 (semireal-sketch, single control, native 1024)
Edit-mode style-transfer LoRA. Successor to 2511-style-transfer-semireal (v4) — same hyperparameters, new dataset: 204 target arts + 204 matching semireal-sketch controls + AI-generated detailed captions (Grok-4.20 vision).
Status: training complete (15 epochs).
Trigger: t01nstyle
Recommended checkpoint
Start with t01nstyle_edit2511_semireal_v5-000013.safetensors (minimum loss at epoch 13: 0.04870). A/B test against epochs +/- 2 by visual quality.
Per-epoch loss
| epoch | loss |
|---|---|
| 1 | 0.05660 |
| 2 | 0.05460 |
| 3 | 0.05220 |
| 4 | 0.05270 |
| 5 | 0.05310 |
| 6 | 0.05260 |
| 7 | 0.05000 |
| 8 | 0.05080 |
| 9 | 0.05200 |
| 10 | 0.05330 |
| 11 | 0.05130 |
| 12 | 0.05200 |
| 13 | 0.04870 (min) |
| 14 | 0.04910 |
| 15 | 0.05170 |
Training
- Tuner: musubi-tuner (kohya-ss),
--model_version edit-2511 --edit_plus - Dataset: 204 art targets x 1 control set (semireal-sketch, native 1024) x 2 captioning modes = 408 effective pairs/epoch
- 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)
- Learning rate: 5e-5 (adamw8bit, constant_with_warmup, 200 warmup steps)
- Timestep sampling: qwen_shift
- Batch size: 4, control_resolution [1024, 1024] (native, no downscale)
- 15 epochs, 1680 total steps
- Hardware: H100 SXM 80GB (bf16, sdpa)
Differences from v4
| v4 | v5 (this run) | |
|---|---|---|
| Targets | 256 (same as v2/v3) | 204 (new, curated) |
| Captions | manual/dual | Grok-4.20 vision auto-captioned |
| Control set | semireal-sketch | semireal-sketch (regenerated) |
| Steps/epoch | 128 | 112 |
| Total steps | 1920 | 1680 |
| Everything else | identical | identical |
Inference
- Pipeline: Qwen-Image-Edit-2511 with control image (semireal sketch)
- LoRA scale: 1.0 to start
- Sampler: FlowMatch / Euler, 25-40 steps, true CFG 4.0
- Resolution: target ~1024^2, control 1024^2 (training resolution)
Dataset sample
See dataset_sample/ folder:
sample_target.jpeg— t01nstyle targetsample_control.jpeg— semireal sketch (control input)sample_caption.txt— detailed caption with trigger
- License
- apache-2.0
- Created
- Jun 3, 2026
- Updated
- Jun 3, 2026