Zaytron40k/Qwen-Image-RenderStyle-LoRA

Qwen-Image (base) LoRA - Render Style (t01nstyle)

A base Qwen-Image (t2i) LoRA that locks in the t01nstyle render aesthetic (shading, linework, skin rendering). Stackable on the Qwen-Image-Edit-2511 pipeline to reinforce style. Trained with DiffSynth-Studio on 337 t01nstyle images.

Trigger: t01nstyle (prepend to the prompt). Trained with dual captions (50% full description + 50% trigger-only) to bind the style to the trigger.

Training config

base	Qwen/Qwen-Image (DiT only, t2i)
rank / lr	32 / 1e-4
epochs x steps	10 x 674 (337 imgs, dual-caption, repeat 1)
resolution	max_pixels 1763584 (~1328^2)
precision	bf16 + gradient checkpointing

Loss

epoch	step	EMA loss min
0	1	0.0478 <- global min
1	1092	0.0654
2	1494	0.0659
3	2602	0.0679
4	2854	0.0661
5	3981	0.0657
6	4560	0.0642
7	5352	0.0610
8	5773	0.0640
9	6179	0.0670

Loss is flat/noisy (normal) - pick the checkpoint by visual eval on WaveSpeed, not by loss.