abide1622/moss-lora-lipbiting-v2

MOSS Image Lip Biting LoRA (v2)

A concept LoRA for MOSS Image that adds a natural lip-biting expression: the upper teeth resting on or gently pressing into the lower lip. It is trained to change the mouth and leave the rest of the image alone.

Usage

• Trigger word: lpb1te. Put it anywhere in the prompt; the examples put it at the end.
• Strength: 1.0–1.2. Below 1.0 the effect is a light tuck; at 1.0–1.2 the bite is clear. With a character LoRA, the composition can start to shift at the top of that range.
• Base models: trained on MOSS Image Base. It also works on SFT-based checkpoints and merges.
• Prompting: describe a relaxed or quiet expression and let the LoRA supply the bite. Prompts that ask for a big smile or an open mouth will fight it.

With character LoRAs: delay the character

On MOSS, a character or identity LoRA running from the first step tends to settle the mouth before this LoRA can, and the bite disappears. The fix is to hold the character LoRA back for the first ~40% of the steps. This LoRA shapes the expression while the composition forms; the character LoRA then takes over the face and keeps the likeness.

In ComfyUI this is a two-pass split with KSamplerAdvanced. At 25 steps:

1. Steps 0–10: model with this LoRA only.
   • add_noise enabled, return_with_leftover_noise enabled
2. Steps 10–25: model with this LoRA plus the character LoRA (at, say, 0.8), continuing from the first pass's latent.
   • add_noise disabled, return_with_leftover_noise disabled

With this split, use this LoRA at 1.0–1.2.

Example prompt

Upper-body portrait photograph of a woman in her mid-twenties with long strawberry-blonde hair in loose, tousled waves, a few strands falling across her cheek, the rest tucked behind one ear. Grey-green eyes, a cute youthful face with delicate features and a light scattering of freckles across her nose and cheeks. Fair, smooth skin; minimal makeup that highlights her natural features. She wears a deep wine-red lace plunge-neck top that frames her collarbones, with a fine gold chain necklace. Relaxed, quiet expression, looking at the camera. She stands against a plain, light-toned studio backdrop. Modern fashion-editorial style. 2/3 dynamic camera angle. Shot on an 85mm lens wide open at f/1.4, razor-thin depth of field, creamy circular bokeh. Warm low-key studio light from the side, golden rays catching her hair, gentle falloff into shadow, golden-amber highlights, deep warm blacks, dim volumetric contrast. Kodak Portra 400 film look, fine organic grain, slight halation around highlights, muted color palette, soft rolled-off highlights. Tactile textures, fabric and skin rendered in fine sharp detail. Dreamy, sensual mood, quiet and private atmosphere. lpb1te

Seed 20261108, 896×1152.

Training

The LoRA was trained with ai-toolkit on MOSS Image Base (bf16). The approach is built around keeping the effect on the mouth:

• Data: captioned portraits showing a lip bite, each with a reviewed mouth mask, in three groups:

  • a larger set of varied faces, repeated 3×
  • a set of especially clear, strong bites, repeated 2×
  • an older set, repeated 1×

  The groups are pooled and shuffled together.

• Masked loss: the mouth mask concentrates learning on the lips and teeth. The rest of the image still contributes at a 0.15 floor.

• Captions: used as written, with the trigger word kept in place and 5% caption dropout.

• MOSS defaults: ai-toolkit's MOSS Image settings.

  • Shifted timestep sampling.
  • The transformer and text encoder are quantised for training.
  • Layer 22 is left out of the LoRA, as ai-toolkit recommends for this architecture.

Setting	Value
Network	LoRA, rank 16, alpha 16
Steps	3,000
Batch size	1
Optimizer	AdamW 8-bit, learning rate 1e-4
Noise schedule	flow matching, shifted timesteps
Resolution	1024, aspect-ratio buckets
Precision	bf16
Hardware	1× NVIDIA H100 80GB

The record of the run is in training/:

• training-settings.json: the training settings, the workflow package and trainer image they ran on, the dataset at an exact commit, and the base model file's SHA-256.
  • Pins the workflow package moss-lora v1 and the trainer image ostris/aitoolkit@sha256:d0b5f729… (ai-toolkit 0.13.41).
• ai-toolkit-config.json: the ai-toolkit config that produces these weights.

Dataset: abide1622/lip-biting-V2 at commit 811aa3bf, subsets main, legacy and strong (94 images).

Images

Strength 1.0, no character LoRA.

Rows: no character LoRA (None), then three different character LoRAs (A, B, C) at strength 0.8, to show how the bite combines with an identity. Columns: LoRA strength 0, 0.6, 0.8, 1.0 and 1.2. Same prompt and seed (20261108) in every column of a row.

Rendered on a MOSS Image SFT-based photoreal merge: 25 steps, er_sde sampler, simple scheduler, CFG 3.5, 896×1152, with the model card's negative prompt. Character LoRAs are held back for the first 40% of the steps (see Usage).

Limitations

• Use it in prompts that show a face. With no face in the scene, the trigger word can end up drawn as text (on a sign, for example).
• The training images are portraits of adult women. The effect works best on similar framing (head-and-shoulders to upper body) and may be weaker on wide shots or other subjects.
• With a character LoRA running from the first step, the bite can disappear; delay the character as described above.

License

CC BY-NC 4.0, following the license of the training dataset: non-commercial use, with attribution. MOSS Image Base itself is Apache-2.0.