A bare-feet concept LoRA for MOSS-Image. Trigger word: feet

Trained on 300 real studio photographs (watermark removal, resolution

bucketing, natural-language captions). It reliably produces the

"both soles facing the camera, toes clearly defined" composition,

and stays clean: without the trigger word it does NOT force feet into

unrelated scenes.

Recommended: er_sde / 25 steps / cfg 3.0 / empty negative / strength 0.9

---

## Tested base models

Works on all 9 MOSS bases I tested. Best results:

| Base | Steps / CFG | Note |

|---|---|---|

| moss_image_sft | 25 / 3.0 | cleanest, official aesthetic |

| Moss-dirtyMoss | 25 / 3.0 | casual snapshot look, 8s |

| Moss-artiusMoss | 10 / 1.0 | fastest (4-6s), built-in turbo LoRA |

| Moss-moodyMossMix | 25 / 3.0 | best skin texture |

| Moss-大师pornmasterMoss | 25 / 3.0 | keeps western adult faces best |

Avoid Moss-动漫miaomiaoHaremMOSS — body-shot composition tends to break.

---

## What it does / does not do

✅ Verified

- Close-up with soles facing camera

- Both feet side by side, toes spread

- Socks (half-off style appears correctly when prompted)

- Full-body shots stay normal when feet is not written

⚠️ Known limitation

- Faces drift toward East-Asian / younger appearance. The training set came

from one demographic, so that trait is learned along with the concept.

This is not a plug-in that only adds feet. To suppress: strength 0.7 +

face-lock prompt caucasian woman, mature face, adult).

- Toenail polish appears only sometimes (56/300 in the dataset)

❌ Don't expect

- Style / lighting change (it's a concept LoRA, not a style LoRA)

- Feet appearing without the trigger word

---

## Training details

| Item | Value |

|---|---|

| Base | MOSS-Image Base (bf16) |

| Trainer | kohya sd-scripts moss_image_train_network.py |

| Dataset | 300 real photos, 4 buckets |

| Steps | 2400 (300 × repeat2 × 4 epochs) |

| Time | 1h20m @ 1.93 s/step on RTX 5090 |

| Network | dim 32 / alpha 32, networks.lora_generic |

| Timestep | truncated_logit_normal (logit_mean -1.09, logit_std 0.95) |

| Text encoder | frozen (official recommendation) |

---

## Compatibility

Requires a ComfyUI build with MOSS-Image support (0.42.0+ tested).

CLIPLoader type must be moss_image. VAE is flux2-vae.

The LoRA is in kohya format and loads directly — no conversion needed.