hanfu.safetensors

632 MB · SafeTensor

Lidmaatschap vereist

<hanfu-flux2>

模型改善了原版的汉服材质，学习了部分汉服款式，优化了亚裔面部的效果（意料之外的好处） 自己从网上扒图片还有收藏小姐姐构成的数据集，不包含名字，可能偏向于本人审美； 本人对汉服不是很了解（只是觉得啥好看就选啥），将近1300张的图片也不可能手标，图片标注用的是grok，提示词工程只能说不尽如人意，但为了省token，标注的不是很好，为了避免拟合人物，没有面部描写，训练的时候中英文都用了，可以直接中文触发； 为了不抑制nsfw内容，在4500步时（汉服元素基本可用了），加了点裸体的素材，把lr对折，继续训练，但平心而论效果一般，只能做到不抑制，需要配合其他lora使用； 尝试加入tang，song，ming的汉服风格（训练的时候加了触发词，可以试试看），但实测很一般； 这是我练的第一个lora，花了我两天的饭钱，从零开始边学边练花了两天，目前算力服务器和大模型中转两个平台的余额也不支持重新再来一遍了； 求求大佬教教我怎么训练lora，怎么搞提示词工程。 Personal Aesthetic Body & Subtle Hanfu Flux LoRA This Flux LoRA is trained on my personal curated dataset — scraped web images + my favorite "miss sister" (aesthetic girl) photos that match my subjective taste in body proportions, poses, and overall vibe. Key dataset & training notes: - No real people/names; captions deliberately avoid any facial descriptions to prevent overfitting to specific individuals. - Captions generated with Grok (kept short to save tokens → basic/generic quality). - Used mixed English + Chinese captions during training for better bilingual prompt handling. - Added limited NSFW material to mildly reduce censorship — it helps prevent outright refusal on explicit prompts, but doesn't deliver high-detail hardcore/genital realism alone. For proper NSFW results, stack with dedicated anatomy/explicit LoRAs (e.g., detail enhancers, pussy fix). - Attempted to infuse Tang/Song/Ming Hanfu styles, but the influence is subtle/weak in tests — hanfu elements appear lightly when prompted strongly. What this LoRA captures best: - Attractive female bodies in my personal aesthetic: slim-thick curves, soft skin texture, seductive/teasing poses, elegant atmospheric feel. - Semi-NSFW or erotic mood shots with decent coherence. - **Notable strength vs. stock Flux.2 Klein:** Significantly improved clothing/fabric texture and material feel (silk, embroidery, flowy hanfu draping looks much more natural and detailed, less "plastic" than base Klein). Trigger word: hanfu, girl LoRA weight: 0.6–0.8 Sampler: Euler simple CFG scale: 2.8 This is a personal "vibe/aesthetic" LoRA — great as a base layer for your own style, but results shine most with good prompt engineering and LoRA stacking. Merge-friendly! NSFW note: For adult/creative use. Generate responsibly. Enjoy the improved fabric realism & happy prompting! ♡ Feedback/remixes welcome.

SHA256

06A985BAAB15727C222E9D5EA8F7A820633582402AB47DC9D63DC5172F946298

ModelTrends.ai model-ID

#26425

Sampler

Euler

Stappen

30

Formaat

928x1392

Formaat

928x1392

Formaat

928x1392