Zaytron40k/Qwen-Image-GenatomyFixer

Qwen-Image (base) LoRA - Anatomy / Genital Fixer (stackable)

A base Qwen-Image (t2i) LoRA that improves explicit anatomy. Designed to be stacked at low strength over an edit LoRA on Qwen-Image-Edit-2511 (the DiT target modules are shared), to fix Qwen's malformed/confused explicit anatomy. Trained with DiffSynth-Studio on 1170 close-up/medium anatomy crops.

Triggers: global n5fw activates the fixer domain; per-part tokens control specific anatomy (e.g. z40peni penis, z41vagn vulva, z42anus anus, z60erct erection, z61flac flaccid, z01vpen vaginal penetration, ...). Full trigger reference: TRIGGERS.md. Blurred dataset samples in dataset_example/.

Stacking usage (on Qwen-Image-Edit-2511)

pipe.load_lora(pipe.dit, "edit_charsheet/epoch-8.safetensors")              # forward edit LoRA @1.0
pipe.load_lora(pipe.dit, "checkpoints/epoch-N.safetensors", alpha=0.3)      # this fixer @low strength
# include n5fw + the relevant z-tokens in the prompt; sweep alpha 0.2/0.3/0.4/0.5

Training config

base	Qwen/Qwen-Image (DiT only, t2i)
rank / lr	64 / 7e-5
epochs x steps	10 x 1404 (1170 imgs, +20% short dual-caption, repeat 1)
resolution	max_pixels 1763584 (~1328^2)
precision	bf16 + gradient checkpointing
captions	natural-language + custom anatomy tokens; global trigger n5fw

Loss

epoch	step	EMA loss min
0	344	0.0513
1	2336	0.0531
2	2997	0.0488
3	4862	0.0531
4	6119	0.0494
5	7741	0.0490
6	9226	0.0502
7	10125	0.0478 <- global min
8	11244	0.0490
9	12945	0.0499

Loss is flat/noisy (normal) - pick the checkpoint by visual eval when stacked, not by loss. Tune the stacking alpha for the best anatomy fix without style/likeness interference.