Weight Slider · MOSS Image
A universal body-mass slider for MOSS Image: slide negative for a leaner build, positive for a fuller one. One LoRA, one knob, no trigger word.
What it does
• Any gender, any age. Trained as a concept slider on a 90-frame mixed dataset (women, men, older adults, varied scenes and crops), so it is gender- and age-agnostic.
• Breast size stays put. The slider is anchor-trained on the bust: body mass moves in the waist, belly, hips, arms, thighs and facial fullness while bust size holds in the working range.
• No side effects. Same seed → same pose, same framing, same colour balance, same face. Only the body changes.
How to use
• Working range −3 … +3. ±1 is a visible, realistic step; −3 reads slim, +3 reads full.
• ±4 … ±5 for exaggeration — at the extremes the bust starts to follow the body.
• Recommended weight 1.0 as the unit; drive strength with the LoRA weight, e.g. <lora:weight_slider_moss:-2>.
• Stacks with character and style LoRAs — load it after them. Any prompt, no trigger. Tested up to 2 MP.
Cover
Same seed and prompt at −5 / 0 / +5, MOSS Image SFT, no other LoRAs — only the slider.
Notes
• Trained with AI-Toolkit (concept slider, rank 4). Dataset fully synthetic, no real people.
• A bust-size slider and a natural/augmented shape slider are in progress.
Description
v1 — guidance 3.0, lr 1e-4, alpha 4, plain "thin / chubby" prompts. Overshot by step 100: ±2 was extreme, composition drifted and the colour cast moved with weight (heavier = warmer, thinner = colder, mean R−B +20 → +3). Cause: a concept slider learns the whole difference between the two prompts, including the cosy/cold prior those words carry.
• v2 — guidance 1.5, lr 5e-5, alpha 1 (paper default), symmetric prompts sharing a neutral tail (natural daylight, neutral colours), bust anchor at 2.0. Cast gone, composition locked — but the male sample's face drifted (stubble, age) and the bust anchor weakened by step 200. Also learned that strength plateaus by ~50–100 steps and later steps only add side effects → 300 steps instead of 800.
• v3 — guidance 0.75, anchor = man's face + woman's bust, anchor 4.0. Everything held, but ±1 read too weak.
• v4 — guidance 1.0, anchor 6.0, poles swapped so + = heavier, samples at ±2. Retrained in parallel as v4b for comparison: slightly stronger, but holds the bust worse than v5.
• v5 — guidance 1.35, anchor 6.0, 300 steps, checkpoint 100: cast within ±2, composition and face identity clean, minimal bust drift. The kitchen ladder (−5 … +5, same seed, 2 MP) confirmed it: working range ±3; from +4 the bust starts following the body.