This LoRA will produce exactly what you ask for; your prompt is just flawed, 99% of the time.

It defaults to a stylized representation of a female nude body. This is a deliberate training preference. If you want saggy tits or other specific variations, you’ll need to guide it yourself through prompting.

If you want the exact same proportions, simply add "proportions, scale, perspective view from the reference image" to your list of maintained features in the prompt. How well it does, is depends on how well you prompt what you want, and what you need preserved without one part overriding another.

Nearly all issues users encounter come down to prompt drift or incomplete preservation instructions. Klein is both an editing and generating model. The moment you add descriptive terms that go beyond pure editing (e.g. "beautiful face", "perfect body", etc.), you introduce potential for creative interpretation.

How that change manifests depends on:

• How clearly you write the prompt

• How well Qwen3 8B understands your logic

• Prompt order and structure

• How effectively you guard the features you want preserved

You can’t have a highly detailed, high-quality LoRA without it generating detail — the key is balancing how strongly you protect the original image versus how much you let the LoRA override. Use strong, explicit preservation language and reference weighting (0.5–0.85) for the most accurate results. Remember: you are asking for something that was not actually fully visible.

Top 3 Reasons This LoRA Becomes Your Go-To:

1. Strong reference fidelity at low weight - Use ~0.5–0.8 and it delivers smoother, more detailed skin while fully preserving the original identity, pose, composition, and body structure. Minimal drift even in multi-pass edits.

2. Progressive high-res training - Started at 1024px, scaled to 1280px (~850 epochs), then 1536px (~1500 epochs) in AI Toolkit for superior detail and upscaling performance.

3. Native 1088x1920 support - Handles tall aspect ratios cleanly and intelligently upscales/enhances your reference images (highly recommended workflow).

NOTES:
-> Strength: 1.0–1.3 (peaks here for perfect nudes in txt2img, img2img, and hybrid workflows).
-> Best for: Photorealistic nudes, body perfection, skin enhancement on Klein 9B.
-> Asian bias note: Dataset leans Asian-influenced, but strong facial locks and proper prompting maintain original identity reliably.

Prompt Guide (Img2Img / Reference Editing):
Editing Prompt (Reference Image) -> Two ways to do this right.
--- --- --- ---
1. Best Practice (Low Drift – Recommended):

Recapture the reference image as a true-to-life authentic optical photograph captured with a Sony A7R IV camera using FE 85mm F1.4 GM lens at f/2.0, 1/250s, ISO 100. Use the provided reference image as visual ground truth for composition, perspective, lighting direction, and spatial relationships. Preserve her body structure, body shape, pose, facial identity, facial features, facial structure, skin texture, skin tone, hairstyle, landmarks, exact position, and background completely. Remove top and bottom garments so the subject is fully nude and fully exposed with detailed body while preserving natural finger grouping and hand anatomy. Apply authentic detailed skin texture featuring organic varied pores, natural skin matte appearance with subtle natural oil with specular highlights. Keep everything else exactly unchanged from the reference image. Only the specified delta is modified.
--- --- --- ---
Potential for drift: LOW
--- --- --- ---
2. Simpler Delta Style (Higher Drift Risk):

Recapture the reference as a true-to-life authentic optical photograph. Remove all garments for full nudity with perfect curvy-fit body. The subject has fake tits with round shape. Apply authentic detailed skin texture.
--- --- --- ---
Potential for drift: HIGH
--- --- --- ---

Pro Tips:

• Always include strong recapture language for identity/pose locking.

• Test weights incrementally.

• Combine with Klein 9B's native strengths for cinematic, high-detail nudes.