My first training try, any feedback means alot! Please comment and share your creation.If you are like me tired of AI faces and would like to see unique faces with true professional lighting and styles, give it a try.Japanese Gravure Photo Style — Krea 2 LoRA A style LoRA that pushes Krea 2 toward the look of Japanese gravure (グラビア) photo books: real-camera skin texture, soft window / studio light, hotel rooms, ryokan, poolside and beach sets, and the typical poses, framing and outfits of that genre (swimwear, lingerie, knitwear, uniforms, yukata…). It is NOT a person LoRA — faces are generic and vary from seed to seed. It is meant as a photographic "base layer" that you stack a character LoRA on top of. 18+ only. Trained exclusively on adult subjects; the dataset was filtered for that before captioning. Please tag your posts accordingly. HOW TO USE - Trigger word: wm_asianmodel — put it at the very start of the prompt. - Weight: 0.8–1.0 alone; 0.6–0.8 when stacked under a character LoRA (load the character LoRA first at 0.8–1.0). - Trained on Krea 2 Raw, used mainly on Krea 2 Turbo and Krea 2 mixes (tested with moodyKrea2Mix).PROMPTING The dataset was captioned in long natural-language paragraphs, not booru tags, so write full sentences in roughly this order: framing → camera height/distance → pose (hands, legs) → expression → hair → outfit and fabric → light direction and quality → environment. The more concrete the sentence, the more the LoRA has to work with.Example: wm_asianmodel, This is a real photograph of a real person. A waist-up shot from slightly above eye level, camera close to her, centered in the frame. She sits on the edge of a hotel bed, one hand resting on the sheets, the other tucking hair behind her ear, looking straight into the lens with a small smile. Shoulder-length black hair. She wears an oversized cream knit sweater slipping off one shoulder and denim shorts. Soft window light from the left, gentle shadows, warm afternoon tone. White sheets, sheer curtains, blurred city view behind her. Natural digital color with accurate, unretouched skin tones and fine skin texture, shot on a full-frame mirrorless camera. TRAINING ≈9,000 gravure photos, each captioned by a vision-language model with a 120–220 word structured description (framing, camera, pose, hands/feet, outfit, fabric, light, environment). The photo-site logo/URL strip was cropped out of every image before training, so the LoRA does not reproduce watermarks or text. ai-toolkit on Krea 2 Raw: rank 32 / alpha 32, lr 1e-4 (adamw8bit), batch 1 × accum 2, 10k steps, flowmatch + linear timesteps, 512/768/1024 buckets, caption dropout 0.05. KNOWN LIMITS - Strong bias toward East Asian women and indoor/resort sets — that is the genre. - Hands and feet are still Krea 2's hands and feet; a detailed pose sentence helps. - It is a style, not an identity: don't expect the same face twice without a character LoRA.Try my workflow for best result <a target="_blank" rel="ugc" href="https://civarchive.com/models/2982673/krea-2-two-pass-refine-workflow-no-upscaler?modelVersionId=3381304">https://civarchive.com/models/2982673/krea-2-two-pass-refine-workflow-no-upscaler?modelVersionId=3381304</a>

Krea2 Japanese photo shoot style

This is my first try of realistic lora trained on rtx pro 6000

wm_asianmodel_krea2.safetensors

v1.0

A well performing none real person character lora, mix from over selected 20000 images. Good with other loras, trigger it with prompt like east asian women etc.

Files with SHA256: 88c8f6f083849df595f4507fd5e7cf17e40cdf9f0f6e0d85835af908581071f3