LoRA (Low-Rank Adaptation of Large Language Models) is a popular and lightweight training technique that significantly reduces the number of trainable parameters. It works by inserting a smaller number of new weights into the model and only these are trained.

2023年10月6日 · To train your LORA effectively, you’ll need an assortment of pictures. Quality and variety are key factors here. You should aim for different facial expressions, clothing styles, types of headshots ( 3/4, 1/2, 1/4, Full-Body Portraits ), angles (not just full frontal headshots), different lighting conditions, various lenses…the more diverse ...

2023年5月22日 · I've been trying to train a LoRA to use my face with about 20 images of myself. I have found many resources and many contradict each other. So I've come here with the hope of clarifying some of my questions.

2023年3月12日 · Train the main lora with the whole dataset and train a separate face lora with pictures of the head/face only (cut and upres). Generate the image using the main lora (face will be somewhat similar but weird), then do inpaint on face using the face lora.

2023年5月10日 · It does a face-detection and inpainting pass after the initial generation (and hires fix), which means your first output already has the face fixed somewhat. And then you can go in to correct smaller details.

2024年3月27日 · Unlock the secrets of AI-driven face swapping with LoRA and Adetailer. Learn to transform images seamlessly in our comprehensive guide.

2024年4月25日 · Discover how to master face swapping in ComfyUI using LoRA with our detailed guide. Achieve seamless, realistic results in your AI photography projects.

尝试优化基础模型中缺乏华人形象的问题，使用 500 张人物环境肖像照，典型职业肖像照和环境照训练，力图覆盖不同性别和年龄段。 使用时无需 Trigger Words，在 SD 1.5 下权重一般不超过 0.8，在 SDXL 1.0 下权重可提高到 1.0。 注意：easynegative 和 ng_deepnegative_v1_75t 两个 Embeddings 会影响此 LoRA 效果，请避免同时使用。 SDXL 1.0 对亚洲人形象的支持，相对于 SD 1.5 已经有了很大的提升。 我仍然针对 SDXL 训练了一个版本，在基础模型表现不佳的情况下 …

Face Detail LORA, the LORA model that I use to generate my best images. Along with this LORA I add other services such as: - My best prompts - Support to configure the LORA model. LORA used for all the images on my website: https://equinox-ai.myshopify.com/

2024年9月21日 · The SameFace Fix Lora is a specialized model designed to counteract the generic faces generated by Flux AI. Unlike traditional negative prompts, applying this Lora with a negative weight tells Flux AI to avoid generating the same repetitive faces, resulting in unique and beautiful outputs.