Breast Size Slider – LoRA for MOSS Image (NSFW, adults only)
One slider, one file, no trigger word. Negative strength makes the chest smaller and flatter; positive strength makes it larger and heavier, with a natural sag. Face, hair, pose and background are meant to stay the same. Made for front-facing and three-quarter pictures of adult women. Not tested on side or back views, so read the limits below.
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
🎯 What it does
Same prompt and seed, MOSS Image SFT, with the LoRA at different strengths. Left to right: -1.5, -1.0, -0.5, 0 (no change), +0.5, +1.0. The size changes step by step; the woman, the room and the lighting stay recognisable.






🚀 Quick start
Base model: MOSS Image (trained and tested on MOSS Image SFT, int8)
Trigger word: none
Strength range that worked in our tests: -1.5 to +1.0
Sweet spots: -1.0 to -1.5 for small and flat, +0.5 to +1.0 for large and heavy
Strength 0 = the picture without the LoRA
Settings used for the examples: SFT, 50 steps, CFG 3, euler / sgm_uniform, shift 5, 896x1344
Prompt tip: ask for a front-facing or slightly turned adult woman with the whole head and face in frame. Say nothing about breast size in the prompt, let the slider decide
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
⚠️ Limits you should know
Do not go above +1.0. At +1.5 we saw over-round, ball-shaped results, blotchy skin tone and, in one case, a swollen body with a visible seam
At +1.0 the hips or belly can look slightly wider, and small props or the view outside a window can change a little
Front-facing and slightly turned pictures work best. When we asked MOSS Image for a side or "profile" view it often produced a back view, and the effect there was less reliable (the face also looked older at high strength)
Tested on 5 prompts with 2 seeds each, all adult women, strengths -1.5 to +1.5. Other styles, clothing, groups or base models were not tested
For adults only. Do not use it for anyone who looks under 18
✨ Features
One small safetensors file, about 8 MB, rank 4, no trigger word
Slider behaviour: strength 0 leaves the picture unchanged, both directions from one file
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
🧪 How it was made
Trained locally with ai-toolkit (concept slider with a region weight on the chest) on 21 pairs of AI-generated adult pictures in which only the chest differs. No photos of real people were used. Checkpoint at step 210. All example images are AI-generated.
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
⭐ Found this useful?
Like if it saved you time
Comment your results — I read every one
Follow for new ComfyUI workflows, all tested on 16 GB VRAM
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
⚖️ Attribution and licence
Base model: MOSS-Image by OpenMOSS-Team, weights released under Apache-2.0 (model card). This LoRA is a modification trained on top of it
Trained with ai-toolkit by Ostris
Training pictures were generated by me; their licences are recorded in my own project notes
Description
v1.0 - first release. Checkpoint step 210, rank 4, trained on 21 chest-only AI image pairs. Recommended strength -1.5 to +1.0 (do not go above +1.0). No trigger word. Tested on MOSS Image SFT, 50 steps, CFG 3, euler / sgm_uniform, shift 5, 896x1344. Best on front-facing and three-quarter views.










