Lovelyzen style LoRA the release measures that look: 16 prompts, 3 prompt languages, 5 strengths and 4 seeds, 960 images, in one interactive page.
The adapter holds the Lovelyzen look across outfits, rooms, cameras and both illustration and photoreal images. It is a style, not a character. The two source pictures are different people, and the matrix includes identity plates to show what crosses over and what does not.
How to use
Base: MOSS-Image-SFT (int8 ConvRot)
Turbo: 8-step DMD LoRA, strength 1.0
Steps: 8
CFG: 1.0
Size: 832×1216
Trigger: optional (
lovelyzen style)Strength: 1.0 is the accepted midpoint.
The tested range is 0.6 to 1.2. Start at 1.0 and move down before moving up.
The trigger is optional. The 960-image matrix was generated with the 16 prompts exactly as written, without lovelyzen style. Add the phrase if you want the look to lock harder; leave it out when the scene prompt already carries the style.
Starter, no trigger:
1girl, solo, adult woman, black dress, standing, plain studio, full body, soft light
Add lovelyzen style in front if you want the trigger. The captions are plain: a trigger phrase then content tags, no artist name, no quality tags. The adapter also follows full scene descriptions.
The 960-image matrix
The release page is the full sweep: 16 prompts × 3 prompt languages × 5 strengths × 4 seeds = 960 images, plus the two source pictures. It is one self-contained HTML file, 35.6 MB, with side-by-side compare, the five-dial ladder, the contact sheet and the exact YAML.
Download: the zip is attached to the release article at civitai.red/articles/37505. The zip also includes trigger-matrix.html, a 192-frame trigger comparison (400 px thumbnails). The page opens on an 18+ gate. Desktop recommended.
What it does well
The look: pale skin, dark hair, soft anime shading and the same presence across outfits and rooms. The black dress, the beach, the bedroom, the garden and the mirror all read as the same style. Tag lists and natural-language scenes both work.
What it does not do
It is not a character LoRA. The two sources are different people; the identity images in the downloadable matrix and article are a stress test, not a guarantee. The matrix includes two failure prompts on purpose. High strengths drift. 1.2 is the top of the tested range; 1.0 is the midpoint.
Training
Two source images and two captions. DoRA, rank 32, alpha 32, attention only, 420 tensors. 300 steps, batch 1, 1 MP, 2e-5 constant, AdamW 8-bit. One RTX 3090, 18 min 50 s. AI Toolkit 0.13.39. Export 67.1 MB. No ffn branches, no text-encoder training, no EMA, no in-process sampling, no SNR weighting. The yaml can be found in the linked article and optional downloads zip.
Permissions
Credit required. Merges allowed. No commercial use. Do not redistribute the file as your own. The page shows Apache 2.0 because the MOSS-Image base model is Apache 2.0; the adapter permissions above are the ones that apply.
Credits
Adapter, training and page: PlayAI. Base model: MOSS-Image-SFT. Training: AI Toolkit 0.13.39. Source pictures and captions: the Lovelyzen dataset.
No affiliation with any character rights holder. All depicted persons are fictional adults of legal age in the jurisdiction where this material is viewed.
Thanks
@_Envy_ for your quick response time, original yaml I took inspiration from, and willingness to answer my questions.
Feedback
This LoRA needs it, I need it. This is how we get better.
Changelog
v1.0 (2026-10-27): first release, with the 960-image matrix and the interactive page.