Liva, an original character of mine.
Trained on Chroma1-HD (ai-toolkit), 50 images, 2000 steps, rank 32.
Trigger: livahh woman
Settings I used on Civitai:
- LoRA strength 0.9 (0.8 also works)
- 35 steps, CFG 4.5
- 832x1216, 1216x832 for wider scenes
- Negative: cartoon, anime, deformed, extra limbs, bad anatomy, blurry
What I've seen it do well:
- Black hair in two long braids down her back, tall build, curvy hips
- Her leather crop top, low-rise leather trousers and knee boots, when you describe them in the prompt
- Dark, moody lighting: moonlight, dim interiors, misty outdoor scenes
- Semi-realistic 3D look. Adding "realistic photo" and putting "3D render, CGI" in the negative pushes it more photographic
Where it struggles:
- Unusual poses (falling, crouching, lying down) often come out stiff or broken
- It holds on to its own trousers, so other clothes don't always stick
- The two braids sometimes merge into one
- Booru-style tags (low-rise pants, from behind) worked better for me than long sentences
Still testing: other outfits, close-up likeness at different angles, scenes with a second person.
Description
Liva, an original character of mine.
Trained on Chroma1-HD (ai-toolkit), 50 images, 2000 steps, rank 32.
Trigger: livahh woman
Settings I used on Civitai:
- LoRA strength 0.9 (0.8 also works)
- 35 steps, CFG 4.5
- 832x1216, 1216x832 for wider scenes
- Negative: cartoon, anime, deformed, extra limbs, bad anatomy, blurry
What I've seen it do well:
- Black hair in two long braids down her back, tall build, curvy hips
- Her leather crop top, low-rise leather trousers and knee boots, when you describe them in the prompt
- Dark, moody lighting: moonlight, dim interiors, misty outdoor scenes
- Semi-realistic 3D look. Adding "realistic photo" and putting "3D render, CGI" in the negative makes it a bit more photographic
Where it struggles:
- Unusual poses (falling, squatting) often come out stiff or broken
- It holds on to its own trousers, so other clothes don't always stick
- The two braids sometimes merge into one
Tip: booru-style tags (low-rise pants, from behind) worked better for me than long sentences.
Still testing: other outfits, close-up likeness at different angles, scenes with a second person.





