A clean, one-dial body weight slider for Qwen-Image 2512.

• Negative strength = slimmer, positive strength = heavier. At −1 she's slim and lean; at +1 she's soft and plus-size, with a fuller belly, arms, thighs and face.

• Everything else stays put. Same face, age, hair, pose, clothes and background across the dial. It was trained with a face anchor so the heavy end doesn't age her.

• No trigger word. Load it and set the strength; it works on any prompt.

• Works across skin tones and ethnicities: it was trained on a balanced, multi-ethnic set of synthetic photos.

Recommended strength: −1.0 to +1.0. The sweet spot is about ±0.8; past ±1.2 the extremes start to exaggerate.

Tested settings: Qwen-Image 2512 (fp8), 20–25 steps, CFG 3–3.5, euler / simple, ModelSamplingAuraFlow shift 3.1. ComfyUI: LoraLoaderModelOnly; negative strengths are fine.

Tips

• It stacks with other LoRAs. Lower it a little if another body LoRA is also active.

• It respects the prompt: on "athletic" or "fit" prompts +1 adds size but keeps the muscle tone, so you get a stronger, thicker build rather than a soft one. Push to about 1.3 for a bigger change there.

• Framing usually holds, but a strong negative strength can pull the camera back to full body. Put the framing ("full body", "waist-up") in the prompt to pin it.

Training notes

Trained with ostris/ai-toolkit's concept-slider trainer on Qwen-Image-2512 (qfloat8): 250 steps, rank 8, lr 2e-4, 768 px. Direction: "slim, toned, lean build" ↔ "heavy, soft, overweight build", with a face anchor to hold age and identity. The images are synthetic (no real people).

Showcase images

• 01-dial-grid.jpg: the same two women at −1, −0.5, 0, +0.5 and +1. Use this as the cover.

• 02–07: single samples at −1, 0 and +1 for each woman.

These are ai-toolkit training samples (768×1152). Before upload, re-render the cover in ComfyUI so it carries generation metadata; that's in the checklist below.